PIZZA看加密

Apple has not even finished swapping CEOs before the new one is already stepping onto the stage to deliver the results. The hardware side has been stockpiling foldable phones for years and now it’s time for them to finally be unveiled. The outgoing CEO also has old debts that need to be settled—especially the one that was repeatedly pushed back and, at the end, effectively outsourced: Siri. In the days leading up to the event, both English and Chinese pre-release drafts are betting on foldable pricing and the quality of Siri’s demonstration. But what I care about is which one will truly change Apple’s financial ledger.

First, Siri. Apple ultimately chose Google’s Gemini as the underlying engine for the new Siri. Industry reactions to this deal have split into two camps. Gene Munster tallied the numbers: to get Siri from “works” to “good,” Apple would still have to spend a large amount of money, and outsourcing is the most financially efficient option. Dan Ives at Wedbush put it more bluntly: this move can turn the more than 1.6 billion iPhones Apple has in use overnight into an AI distribution channel. On the other hand, skeptics have a point too. Ming-Chi Kuo has noted that the new Siri must be better than the Gemini core itself to count, and being delayed for so long only means it’s catching up—it won’t persuade anyone.

I’m on the first side. Apple’s moat has long been in distribution and the end devices, not the model itself. One user won’t switch phones just because the voice assistant’s engine changed, but every year they must decide whether the device they already have can last another year. So whether Siri’s demo looks impressive tomorrow won’t have as much marginal impact on valuation as the market thinks.

What will move the ledger is gross margin. Apple’s June-quarter results were actually among the most solid in the big group. Revenue came in at $109.4 billion, up 16.4% year over year. iPhone alone rose 21.7%, and gross margin stood at 50.1%. For a company being called behind on AI, there’s no sign of any deterioration in its hardware sales ability.

But management’s own guidance for the next quarter looks less appealing. The gross margin range was compressed to 47%–48%, and the company explicitly said supply constraints are expected to worsen. On the cost side, things are indeed changing. This year’s memory price hikes are an industry-wide event. Per TrendForce, in the second quarter Apple’s NAND contract prices rose about 70%–80% quarter over quarter. The Android camp is hit even harder—nearing a doubling. Institutions have repeatedly lowered their shipment forecasts for global smartphones this year. What phone makers generally do is raise prices and cut specs at the same time. Over the past decade-plus, the tradition has been to trickle flagship specs down to mid- and low-tier models; this year, that trend starts to reverse.

Foldables hit that cost environment perfectly. Ming-Chi Kuo’s price range is $2,300 to $2,500, and the production capacity target this year has been lifted to around 10 million units. Erik Woodring at Morgan Stanley called this the most important product launch since the iPhone X era. On the revenue statement, it will look good. But 10 million units compared with the iPhone’s annual shipment volume looks more like a high-price, narrow product line. Hinges and foldable panels are already expensive; add memory price increases on top and it can lift revenue while also dragging down gross margin. In the financial statements, this kind of product tends to split: the line for revenue looks great, while the gross margin line trails.

That’s where the disagreement comes from. The bullish side treats foldables as the starting point for a new category and is willing to restart valuation based on the replacement cycle; Morgan Stanley’s target price is as high as $360. The bearish side treats it as simply a more expensive configuration—if it won’t sell, it won’t sell. Until now, Dan Ives’s colleague at DA Davidson has held a $270 target. Same company, same product—yet the two sides’ room to move differs by one-third.

Tomorrow, the most important thing to watch is the pricing page. It’s management’s public answer to cost pressures. The starting price lands at the upper end of Kuo’s range, and it doesn’t come with large-scale installment plans or trade-in subsidies, suggesting Apple has the confidence to shift costs to consumers. If it’s priced near the lower end—or if shipments are pushed to later in the year, as Kuo suggested—then it’s prioritizing volume while absorbing the costs itself. In that second scenario, the gross margin guidance for the December quarter will look worse than the foldable’s product impact itself. The event day won’t include any gross margin numbers, but pricing and the shipping cadence have already written the answer into the plan. The rest will be confirmed during the earnings call.

The biggest counterargument to my view also comes from memory. In terms of procurement scale and contract terms, Apple is taking less of a hit than Android during this round of price increases. When costs rise, the Android hardware gross margins—already thinner—cannot take it. If a price war really breaks out, that would actually be an opportunity for Apple. Plus, service revenue is still growing at double digits. Its gross margin is far higher than hardware’s; as its share rises, it can offset part of the hardware pressure. If these two things work, gross margin might not fall to the low end of the guidance.

The price has already moved. The $AAPLB spot price is currently $318.92. Even more telling is the timing: on the day the new boss was appointed, the stock surged, reaching an intraday high of $330.66 on September 3; afterward, it retreated back to where it is now. #苹果发布会 hasn’t opened yet, and expectations have already leaked some in advance.

Tomorrow, the truly valuable information is hidden in the pricing page and the gross margin range for the December quarter. You can look at those two places and then decide how to view this stock. The parameter list on the product page can be left for reviewers to argue over.

The pricing for that September meeting shifted completely in the other direction after last Friday. They were originally debating whether the Fed would cut rates, but now the discussion is whether it will raise rates instead. The employment data far exceeded expectations. There is PPI on Thursday and CPI on Friday—those two data releases directly moved ahead of the meeting. This week has no meaningful earnings reports, so interest rates are the only main storyline.

On the interest-rate futures side, the probability of a September rate hike is currently hovering around 60%. Today is the U.S. Labor Day holiday, and the quotes on $CRWVB are basically just carrying forward last Friday’s move—there’s nothing to see there. In the meantime, I’ll take advantage of the lull to go back and review CoreWeave’s accounts again.

Its predecessor was an Ethereum-mining operation. Today, its business is building clusters out of Nvidia GPUs and renting them out. On its customer list are OpenAI, Meta, Anthropic, and Google. As of the end of June, contract backlog is roughly $104 billion. On the demand side, there’s nothing much to argue about.

The problem is how those cards were bought back.

At the end of the second quarter, its balance sheet shows total debt with recourse and without recourse of about $35.1 billion. In the same quarter, net interest expense was $640 million. In the company’s guidance for the third quarter, interest expense is expected to rise to between $860 million and $940 million, while adjusted operating profit in the same guidance is between $200 million and $260 million. Interest alone is roughly three or four times operating profit—this ratio isn’t just especially ugly in one quarter; it has been like this all along.

If you extend it to the full year, the ratio becomes even more extreme. Management’s capital expenditure guidance is $35–39 billion, while revenue guidance for the same year is only a bit more than one-third of that—everything in the gap has to be covered by financing. The interest CoreWeave pays each quarter is real cash—changes in interest rates flow directly into its income statement, without having to wait for valuation-model transmission.

If rate hikes land on its head, my original thought was about the valuation layer. But the announcement of the new $2.6 billion loan on August 10 changed that—there’s harder stuff baked into the liability structure.

That was a $2.6 billion delayed-draw term loan priced at SOFR plus 550 basis points, with a maturity of about five years. The announcement itself states that the underlying customer contracts have an average remaining term of about three years.

The assets mature in three years, but the liabilities are not fully repaid until five years. The two-year gap in the middle has to be filled by renewals or new customers. When the market is hot, this isn’t a problem—the cards are scarce and renewals queue up. But once the demand rhythm changes, those two years become an open exposure.

How the company explains this money is one thing, and how you read the terms is another. The CEO’s explanation on the earnings call was that this structure allows the company to serve enterprise customers that only want to sign for two to three years—sounds like it opens up a new market. Truist analyst Arvind Ramnani gave bullish reasons at the end of August, and his conclusion lands precisely on long-term contracts. He thinks the pricing increase will first pass through into customer contracts, and cost increases on Nvidia’s side won’t show up until the machines that ship in early next year. Over the intervening months, the contribution margin from long-term deals would be lifted clearly.

The company is opening the door to short-term contracts, yet the bullish logic is built on the margin from long-term contracts. Both can be true at the same time, but they point to two different businesses.

How exactly this rate line will hit, the CFO has already given the calibration. He said that over the past year, the weighted average cost of debt was lowered by about 300 basis points. Based on the debt balance at quarter-end, that saves roughly $1.1 billion in interest per year. Those 300 basis points came from improving credit conditions and opening up financing channels—not from the benchmark interest rate. If you flip that same logic and apply it the other way: for the same debt size, if the benchmark rate rises by 25 basis points, the additional interest paid in a year comes to nearly $100 million. After two or three hikes, a large chunk of what management worked hard to save would be eaten away.

So if this round truly moves into a rate-hike path, what the market needs to reprice first is CoreWeave’s ability to refinance and the financing-cost curve; the P/E multiple comes later. On the demand side, I don’t doubt it—those four biggest model companies have signed, and you can’t draw those orders out of thin air. What I question is whether, in an environment where borrowing gets more expensive, the three-year-to-five-year gap might shift from a technical arrangement into a real problem.

Evidence to overturn that view should come this week. If Thursday’s PPI and Friday’s CPI both cool off, and the rate-hike bets get cut back to below 40%, then the stress-test calculations above would be discounted overall. Conversely, if both data points come in hot, only next week’s meeting will put this logic on the table.

For a heavy-asset AI company, that contract backlog number is so large that it can lull people into complacency. When the next earnings report comes—#CoreWeave —you can check whether the gap between interest expense and adjusted operating profit continues to widen. How much that gap has expanded, more than how much the backlog has grown, is what better shows how far the company can hold on into which year.

Around HYPE setting a new all-time high, the explanations on the market have been highly consistent. Since it kept climbing right around the monthly vesting date, it must have been protocol buybacks absorbing the sell pressure. That sounds plausible, but there is one link in the middle that nobody checked: whether the coins written into the vesting schedule ever actually left the custody address.

Hyperliquid’s own supply interface is public, and the answer is there. The address that holds the core contributors’ vested share received 238 million tokens at genesis, and today it reads 241.16 million, a bit more than at launch. The extra amount is staking yield, because the entire balance at that address is delegated, and the available balance is zero. The 9.92 million tokens that vest each month have never become part of the circulating supply on the protocol’s own ledger. That number appears on the calendar every month, but not in the market. Vesting expiration only removes one restriction; for the coins to actually reach the market, someone still has to actively un-delegate, withdraw, and place sell orders. On-chain, those are three separate actions, and each one has to be deliberately executed.

The price picture also should not be overread. The all-time high was touched on the evening of September 6 at 89.66, but the close that day did not hold above it, and spot today is quoted at 86.39 on OKX. Touching an intraday high and holding above it are two different things; treating a wick as trend confirmation turns a single candle shadow into a conclusion.

Let’s do one more layer of arithmetic. Suppose those tokens were actually claimed in a given month and dumped all at once the same day. At current prices, the notional amount would be around $860 million. Over the same time frame, Hyperliquid’s protocol revenue over the past 30 days was $55.35 million, and that is the full ammunition of the assistance fund for buying back HYPE. Using one month’s revenue to absorb one month’s nominal unlock amount would only cover a small fraction. So attributing this new high to buybacks is looking in the wrong direction. Right now, HYPE is being priced off the 299 million tokens that are actually circulating; the vesting schedule numbers never entered that pool.

The real point worth scrutinizing is the revenue line. OAK Research’s Lilian Aliaga compiled a set of figures in late August. Quarterly protocol revenue fell from $357 million in Q3 2025 to $202 million in Q2 this year, and the assistance fund’s buyback amount was cut roughly in half as well. She sees this as an active choice: Hyperliquid is giving more and more fees to developers building products on top of it, trading revenue for activity and market share.

I agree with her on the direction, but I would not stop the risk analysis at revenue. Declining revenue is a slow variable; there are signs before the quarterly report comes out, and readers have time to react. The fast variable is the custody address balance. The 241 million tokens there are equal to 80% of the circulating supply. Although unstaking takes time, the queue is only a few days. If nobody is claiming today, that is a holder’s deliberate choice after doing the math; the contract has not welded these tokens shut. The calculation is actually straightforward. Tokens left staked keep earning yield; withdrawing them means giving up that yield in exchange for an uncertain sale price. As long as HYPE keeps trending upward, staying put is the most profitable move. But that logic depends on the price trend, and trends change; once they do, the answer the same group of people calculates will change too.

The assistance fund is also a two-sided story. Its 47.04 million HYPE were accumulated at an average cost of $27.22, a position close to one-sixth of circulating supply, and the protocol itself is the largest single holder in that pool. When the market rises, this acts as a thick buffer; when revenue falls, it becomes a buyer that is forced to slow down more and more, and everyone knows it cannot keep buying as aggressively.

The vesting story around #Hyperliquid is a false issue right now, but it can turn into a real one at any moment, and the trigger condition is clearly identifiable. If the delegated balance at the custody address starts to decline month by month, that means contributors have begun cashing out, and my interpretation above would fail immediately. If protocol revenue continues to fall along the slope seen in Q2, then buyback support will be reduced to little more than a narrative. Checking these two numbers every month is more useful than staring at the vesting calendar.

Conversely, don’t overstate the risk either. For staked coins to come out, they have to be un-delegated first, and that action is visible on-chain. They won’t suddenly dump out of nowhere at some unannounced dawn.

Next time you see a headline saying some token unlocks billions worth on a given day, first take a look at the project’s own supply interface. How much remains in the custody address? Is it fully delegated? You can check it in two minutes. The gap between nominal unlock amounts and actual circulating supply increases is often an order of magnitude or more, and $HYPE is just the clearest example of that gap right now.

GameStop will report its full second-quarter earnings after the close on Tuesday. Its core business is selling games and collectibles, but the biggest asset on its balance sheet is an eBay stake. This spring, eBay’s board rejected its takeover proposal, calling it neither credible nor attractive. Ryan Cohen did not stop after being turned down; instead, he kept buying, and now he is eBay’s most troublesome shareholder.

The preliminary results at the end of August already told us the big picture. Net profit came in at $290 million to $310 million, up from $168.6 million a year earlier, almost doubling; revenue was $780 million to $800 million, down from $972.2 million a year earlier, a drop of nearly 20%. Profits doubled, but revenue shrank. When those two numbers sit on the same statement, the first question is where the extra money came from.

The announcement itself made it clear. Derivatives and equity investments in eBay generated about $238 million in net gains, partly offset by roughly $75 million in losses from digital assets and related receivables. The in-and-out from those two items accounted for a large share of quarterly net income. Neither was earned from selling merchandise.

I checked the $75 million myself. GameStop held 4,710 BTC, and its second quarter ended on August 1. BTC closed at $78,687 on May 2 and at $62,823 on August 1, a drop of more than $15,000 per coin. Multiply that by 4,710 coins and you get about $74.7 million. That almost exactly matches the roughly $75 million in the announcement. That loss had nothing to do with how well games sold; it was simply the price of bitcoin moving down over those three months.

There is an easy-to-misread detail here. Of those 4,710 coins, 4,709 were pledged as collateral at Coinbase, which had the right to rehypothecate them. Under U.S. accounting rules, those coins have to be removed from the balance sheet and reclassified as digital-asset-related receivables. So the line in the financial statements is a receivable, not a pile of coins. The company also wrote covered calls against this position, and the premium and exercise outcomes are mixed into the same line. The accounting check above matches so closely because the direction is right, but it is not a perfect identity. If Tuesday’s full report shows the number of coins changed, this reconciliation will have to be redone from scratch.

What is being obscured by these two mark-to-market swings is the operating profit line: $150 million to $170 million, up from $66.4 million a year earlier. Revenue fell 20%, yet the money made from selling products more than doubled. Last quarter, its gross margin was already clearly improving, and collectibles had become the largest revenue category for the first time; that trend is still at work.

Steve Eisman said in April that GameStop’s profit improvement was driven by cost cutting, and that expecting it to transform through acquisitions was a fantasy. The cost-cutting part no longer holds up very well; gross margin improved because the mix of what it sold changed. But he also said this was still a shrinking business, and the second-quarter revenue backs him up.

The eBay story needs a few words too. GameStop initially built the exposure through derivatives, then in the second quarter turned it into direct share ownership, causing cash and marketable securities to fall from more than $8 billion a year earlier to just over $5 billion. After the takeover proposal was rejected, Cohen tried to lower the threshold for calling a special shareholder meeting. That proposal was rejected at the June shareholder meeting, with opposition votes clearly exceeding support. Governance went nowhere, so the remaining option is to go around the board and make an offer directly to eBay shareholders. He has publicly said he will not stop and will not walk away.

What concerns me more is that the direction of these two positions has already flipped. On August 1, the 43.4 million eBay shares had a carrying value of $4.947 billion, or about $114 per share. Last Friday, eBay closed at $103.41. On a static basis, that position was worth more than $400 million less than at quarter-end, which is larger than the newly confirmed $238 million gain. At the same time, BTC was back at $79,896, above the August 1 level, so the digital-assets line is now in the money. Last quarter stocks made money and coins lost money; so far this quarter, it is exactly the reverse.

That position also carries another layer of trouble. Nearly 10% of the equity is being used as pressure, not as something to be freely sold at any time. If he really reduced the stake, the price would likely fall first, and the market would immediately read it as Cohen surrendering, which would also kill the acquisition storyline. The mark-to-market gains and losses come and go on the statements every quarter, but whether they are ever realized depends on whether eBay is ultimately taken over. Reading this report as if it were just a portfolio of securities that can be sold at any time would overstate its certainty.

A retailer’s quarterly profit statement is now being determined by two mark-to-market prices, and earnings day is just a snapshot of them. Whether the snapshot looks good has already become a separate question from how well the company actually performed over the quarter.

On Tuesday, I will first look at operating profit and gross margin, because those are the only parts of the report that say anything about the business itself. Then I will check whether collectibles continue to account for a larger share of revenue, and exactly where the 20% decline came from. If gross margin falls back on Tuesday and collectibles stop growing as a share of revenue, then the operating-profit improvement in the first half of the year looks more like a comparison boosted by a weak base last year, and Eisman’s other point will start to hold again. As for whether the eBay takeover bid is still on the table, management probably cannot avoid that question. If you own $GMEB , or just want to understand how $BTC gets recorded once it is placed on the balance sheet of a #美股 -listed company, it is worth taking a look at this full report.

The people standing in the way of the crypto industry in Washington this week are not on the Democratic side. Missouri’s Josh Hawley and Kansas’s Jerry Moran have both said that if the stablecoin rewards language is not changed, they will vote no. In the same camp, Oklahoma’s Lankford and South Dakota’s Rounds have also put the concern about community bank deposit outflows on the table. These lawmakers are from the same party as the bill’s backers.

The Senate has set the vote for 2:15 p.m. Eastern Time on September 15. What is being voted on that day is whether the bill can be brought to the floor for debate; the bill itself will not yet be up for passage, and the threshold is also 60 votes. Republicans hold 53 seats, so by the simplest math they would need at least seven Democrats to reach 60. But if two or three of their own defect first, the number of Democrats needed rises to more than ten.

CLARITY is trying to settle who regulates a token. The bill gives the CFTC authority over spot digital commodity markets, while the part deemed to be securities remains with the SEC. The House has already passed it, and it is stuck in the Senate. After a year of talks, three issues remain unresolved: conflict-of-interest rules for government officials and their families holding crypto assets, the language on enforcement and anti-money-laundering, and the one closest to users: whether stablecoins can pay yields at all.

That last issue has recently changed in nature. Tillis and Alsobrooks worked out a compromise in the spring: rewards economically equivalent to bank deposit interest would be banned, while rewards tied to usage would be allowed. If you use it to pay or trade, the platform can give you rebates; if the money just sits in the account, the platform cannot calculate yield on a daily basis. The banking industry rejects that carve-out. The Bank Policy Institute and the American Bankers Association say that as long as stablecoins can pay any form of yield, deposits will flow out of the regulated banking system, with community banks hit first. Coinbase takes the opposite view, arguing that rewards programs are a competitive necessity and are different from deposit interest. Lummis is on the side of allowing them.

So Republican senators from agricultural states ended up on the banks’ side, clashing with their own party leadership. The conflict between crypto and banking has already been rewritten on #CLARITY法案 as a conflict inside the Republican Party, and the vote gap has never been only on the other side.

The Democratic side has not balanced its books either. A seven-senator joint statement at the end of July said it plainly: the text Republicans have proposed is not enough yet, and ethics, consumer protection, illicit finance, conflicts of interest, and market integrity all need to be strengthened. The signers were Cortez Masto, Alsobrooks, Booker, Gallego, Hickenlooper, Warner, and Warnock. Those seven happen to be the very bloc most likely to break ranks. In late August, Gillibrand made her position final: without an enforceable ban on profit for sitting officials, she will not vote for it.

Prediction markets are more informative than headlines. On Polymarket, the contract on the bill being signed into law this year was at 14.5% this morning; in February it was above 80%, and it has been sliding for seven months. More telling is the adjacent set of contracts broken down by vote total: the contract for the final vote to get more than 50 votes was at 28%, and the one for more than 60 votes was at 23.5%, only a little more than four percentage points apart. Bettors think that if the bill really makes it to a final vote and gets a simple majority, it has basically already found its 60 votes. It will not die in the tally; it will die before it ever reaches the floor.

On the same page there is also a set of contracts betting on whether a particular senator will vote yes. Hawley is at 18.5%, Moran at 14.5%. Those numbers should be discounted, because the settlement rules say that if there is no final vote this year, everyone is settled as not having voted yes, which drags all prices down a layer based on whether the vote happens at all. What can be read is the relative order. In that order, Democrats Warnock and Andy Kim are both around 40%, well above Hawley and Moran. Bettors judge that this bill will lose more Republican votes internally than Democratic votes.

Galaxy’s Alex Thorn cut the odds of enactment this year to 10% in mid-August, arguing that after the Senate reconvenes on September 14, there are only two or three usable weeks left and neither the ethics issue nor the banking issue has been resolved. Armstrong said he expects support from more than 60 votes. That comes from the CEO of the biggest beneficiary, so it can be treated as an industry line, not a vote count. I lean toward Thorn’s view, based on the structure shown by the contract breakdown above: the bottleneck is whether the negotiation can be closed, not persuading Democrats.

$BTC was at $79,929 this morning and barely moved over the course of the day. Over the past seven months, the odds on this bill have fallen from 80% to 15%, and $BTC has not collapsed along with them; the market had already priced in failure. If the September 15 vote falls short, the price is unlikely to make much of a move. What has not been priced in is the other side.

There is only one signal that could overturn the judgment above: a new version of the stablecoin rewards language, followed by one of Hawley, Moran, Lankford, or Rounds publicly changing his position. If that happens, the Polymarket line will move first, and the vote will just be a formality. The Senate reconvenes on September 14, voting starts on the afternoon of the 15th, and there is only a little more than one day in between; who gives way during that window will tell you the outcome sooner than the roll call on the day itself.

The FOMC rate decision comes out on the afternoon of September 16, right after this procedural vote. If the bill fails on the 15th, the next day’s macro news can wash it away within hours, and the week’s crypto policy narrative will be trailing the Fed.

Over the next week, watch the public statements of those Republican senators and whether a new draft of the stablecoin rewards language appears. This line will tell you more clearly than Washington’s upbeat spin how far the bill has actually progressed.

After the clinical data for the cancer vaccine came out, Moderna's management immediately went to the bond market and sold a batch of convertible notes. The notes pay no interest, and the conversion price was set well above the stock price at the time. The terms for the institutions willing to put up the money were simple: the stock would have to rise much further before the paper they bought would start to be worth anything.

The company behind $MRNAB had just posted its biggest single-day gain in history, then turned around and financed itself this way. Management's attitude is already written into the action. At the current price, selling options is more attractive than selling stock.

On August 19, Merck and Moderna announced the results of INTerpath-001. The personalized neoantigen therapy intismeran autogene, combined with Keytruda, was used in patients with high-risk melanoma after complete resection. The primary endpoint of recurrence-free survival was met, and the key secondary endpoint of distant metastasis-free survival was also met. mRNA #癌症疫苗 achieved a phase 3 positive result for the first time, and it was also the first time for the personalized neoantigen approach. On the day of the news, Moderna's common stock jumped from $62.96 to $174.38, its biggest single-day gain since listing.

Only the conclusion was announced. The two companies did not provide a hazard ratio, nor did they give any specific efficacy figures. They only said the results were statistically significant and clinically meaningful, with detailed data to be presented later at a scientific conference, while overall survival is still being followed. The money that rushed in on August 19 was buying a qualitative statement.

And it was buying far more than melanoma. The platform also has nine other trials under way, covering lung cancer, bladder cancer, kidney cancer, pancreatic cancer, and gastric cancer. Success in one tumor type at phase 3 was treated as a pass for the whole pipeline. Simon Baker of Rothschild & Co Redburn downgraded the stock from neutral to sell on September 3, while sharply raising the target price to $81. He said the data were unquestionably good, but the stock's reaction implied the therapy would work across nearly all tumor types, even though there is almost no data for those cancers right now.

Sell-side analysts split into two camps. Myles Minter of William Blair ran the numbers. For melanoma alone, on a 50-50 split with Merck, Moderna's annual peak sales could reach $5.4 billion. The opposing view focused more on the mechanism. Daina Graybosch of Leerink Partners pointed out that melanoma is already the most suitable cancer for a vaccine approach. Extrapolating success here to other cancers involves several additional steps, and the cost of individualized manufacturing is still very real. Cory Kasimov of Evercore ISI noted that the risk-reduction benchmark management used this time was a step back from the phase 2 result.

The phase 2 numbers are public. Follow-up data from KEYNOTE-942 showed the combination therapy reduced the risk of recurrence or death by 49%. With phase 3 scaling the sample to more than one thousand patients, no one knows where the hazard ratio will land today.

Back to the balance sheet. Over the past twelve months, Moderna's revenue of $2.23 billion is still declining, net loss is $3.15 billion, and market cap is $58.1 billion. This combination cannot be explained by conventional valuation methods; what is holding it up is confidence in a pipeline that has not yet disclosed the numbers.

The cash line is clearer. At the end of June, cash and investments stood at $6.9 billion, down from three months earlier, and the company's own year-end guidance is for that to fall to around $5 billion. At that burn rate, without financing, the remaining runway would only be two to three years.

The $3 billion zero-coupon convertible note solved that problem in one shot. The private placement was completed on September 1, with maturity in 2032 and a conversion price of about $210.58. An additional capped call was also put in place, raising the effective dilution threshold another notch. The cost of funds is zero; the price paid is giving up part of future upside.

That $210.58 figure says more than the financing size does. Management was willing to sell future upside at this level, which shows they are at least comfortable with the current stock price. Not paying interest is the other side of the same signal: institutions were willing to give up all coupon income because the stock's current volatility made the embedded option value high enough that interest was unnecessary. And that volatility comes from a risk ratio no one has seen yet.

This stock is currently in a state that cannot be priced, which is a different issue from whether it is expensive. Baker's target price of $81 is not one I really buy. He himself admits the data are good; the source of the $81 number is the valuation model, and it has little to do with this readout. On a biotech stock waiting for a key data readout, assigning a target price precise to the nearest dollar from a valuation model is not convincing.

Whether the view changes depends on when that hazard ratio is released and where it lands. If the phase 3 hazard ratio is close to the phase 2 level, then doubts about extrapolation lose their footing, and the $5.4 billion peak-sales assumption gets a foundation. If it is clearly lower, near the benchmark Kasimov described, the drug will likely still win approval, but the $58 billion-plus market cap will lose its anchor.

That judgment could also be wrong. If William Blair's view proves right, melanoma alone could support a substantial portion of today's valuation, and the other eight tumor types would amount to free options. Merck is covering half the development cost and providing Keytruda along with its entire commercial infrastructure, which is a real burden reduction for a company that is still losing money. Once the financing is in the bank, balance-sheet risk is much lower than it was a month ago, and that alone should lift the valuation of the whole pipeline.

Before the hazard ratio is published, the price is mainly supported by confidence. $MRNAB spot is now quoted at $144.68, down 5.74% in 24 hours, and the common stock has also been trending lower this week. If you want to follow this line, watch for when the companies announce which conference will host the data presentation, and when the first readout outside melanoma in the nine trials comes out. Those two things matter more than the daily price swings in determining whether the valuation can hold.

Goldman Sachs, Citigroup, Bank of America, UBS, Deutsche Bank, and Mitsubishi UFJ—this group of names gathered into the same announcement, saying they plan to co-found a company and issue a dollar stablecoin. In the past couple of days, most Chinese-language posts have stalled at translating the list of names and adding a line about traditional finance making a comeback. The roster is indeed impressive, but it doesn’t answer the more important question: once this money actually comes in, which side does this stablecoin business end up being passive on?

Let’s get the facts straight. A total of 21 institutions signed the letter of intent. In North America, there are also Wells Fargo, Toronto-Dominion (TD), Scotiabank, PNC, and First Capital, plus two asset managers, Fidelity and WisdomTree. Europe includes Santander, BBVA, Crédit Agricole, Lloyd’s, Rabobank, and Commerzbank. Africa is Standard Bank, and the Middle East is Sirius.

The company hasn’t been named and hasn’t been formally established yet. The announcement says it’s intended to be set up, and that delivery is subject to conditions. The entity plans to be built in the second half of this year, with the token targeted to launch in the first half of next year—starting with the dollar, and then expanding to the euro and other G7 currencies. Nothing has happened yet; they’ve simply written down when it will happen.

The issuer’s revenue comes only from the reserves side: collecting users’ dollars and buying short-term treasuries to earn interest. The hard part has always been the distribution side—how to get users to let their money sit in your coin instead of someone else’s. The GENIUS Act rewrites the relationship between the two ends: the issuer is not allowed to pay you, in any form, any interest or yield merely because you hold that stablecoin; cash, tokens, and other consideration are all included. Once this lands, the path of grabbing customers by offering higher yield is effectively sealed. If users hold anyone’s coin, the yield is zero. The only thing left to compare is who occupies the position where users keep their deposits.

Circle’s earnings report describes it more directly than any analysis. In Q2, its total revenue plus reserve income totaled $701 million, with reserve interest alone making up $668 million—so it claims the company has only one item of income, which isn’t an overstatement. In the same quarter, distribution and trading costs were $412 million, with the bulk paid to Coinbase. For every dollar Circle earns from reserve interest, more than half has to go to the party that helps it secure user custody. Since the law doesn’t let it pay users, it can only route the money through channels.

By the end of June, that group already moved with this logic. Stripe, Visa, Mastercard, Coinbase, BlackRock, and more than 140 other companies came together to form Open USD. The mechanism explicitly states that the vast majority of reserve earnings are returned to growth partners that help drive the platform’s growth, while it keeps only a small management fee. The payment Circle makes to Coinbase was directly turned into product design.

At this point, the banks’ targeted position becomes clear. They don’t need to win anyone on yield. They already sit in places like corporate accounts, cross-border settlement, and correspondent bank clearing. Customers’ dollars were already in their hands. #稳定币 is more like giving existing channels another layer of settlement rails—it isn’t about acquiring customers from scratch.

The rate side, on the other hand, is actually a tailwind. The federal funds rate is still above three percent, and in September’s meeting the market is even pricing in rate hikes. Every dollar sitting in reserves now earns more than at this time last year. The issuer’s trouble has never been the level of interest rates; it’s always been whether the money will stay with it.

USDT’s current circulating supply is roughly $183.3 billion, and over the past six months it’s been basically flat. $USDC is roughly $73.8 billion—back in March this year it was at higher levels, and then it shrank steadily, bottoming out in early August, only recently coming back a bit. In early August, Morgan Stanley downgraded Circle from Neutral to Underweight, cutting the target price to $38. The reason was how deeply USDC contraction exposes Circle’s reliance on reserve income.

My view is that the bank group’s move basically doesn’t overlap with USDT. USDT is positioned where emerging markets use it—using cash and doing over-the-counter settlement—supporting most of the exchange’s quotes. A bank coin, compliant-first and aimed at institutional wholesale settlement, cannot squeeze into these places in 2027. What’s being targeted is USDC. It competes with this bank coin for the same customer: institutional dollars within the U.S. regulatory framework. That customer cares about compliance and also about who holds the account—both are the bank’s home turf.

The most fragile part of this scenario is timing. The body only gets established in the second half of this year. The token has to wait until the first half of next year. Even the OCC implementation details are delayed until November for finalization, and after that there will be a rollout window. For a joint venture company formed by 21 shareholders, how slowly decisions get made is something anyone who has run cross-border projects knows well. If one shareholder drops out midstream or they revise the wording, the schedule gets pushed back—and in the meantime Circle can swap several rounds of channel partners.

It’s also possible I’ve got the direction backwards. If, at the end of the day, this coin is only used for wholesale settlement among banks and never truly issued to end users, then it wouldn’t overlap with USDC’s customers, and everything above would not hold. The announcement also mentions two use cases—wholesale transfers for institutions and retail payments. Which level they actually reach will only be known once the product comes out.

There’s one more variable on the rules side, and it favors banks. Banks are currently lobbying regulators to expand the interest-payment ban from issuers to affiliates and exchanges, closing the loophole that effectively routes interest back to users via channels. If this gets written into the final rules, the hardest hit would be the USDC incentive scheme Coinbase runs. But the bank group doesn’t rely on returning interest to acquire customers in the first place.

Over the next few months, this likely won’t leave much of a footprint in market pricing. Structural changes at the beginning are like this—if you really want to track it, watch USDC’s circulating supply. Circle’s own transparency page updates weekly; whether supply is shrinking or growing is more honest than any interpretation. As for the joint venture, you can see whether it truly gets registered by year-end, whether it has a name, and whether any shareholders exit along the way. The day the name is announced is when this goes from being a press release into a real company.

On the day Cook handed over, the market didn’t give Apple any special courtesies. Global bond yields surged, oil prices climbed, and the Nasdaq was all green. Against that backdrop, Apple itself carved out an upward line, turning into one of the few large-cap tech names that finished the day in the red.

What’s even more thought-provoking than the upside is what kind of company the board chose to hand to its next chairman at this point.

First, the person. It began with Timbas in product design, and he rose all the way to senior vice president of hardware engineering. He’s had a hand in every major product line—iPad, AirPods, and Apple Watch. After spending more than twenty years at Apple, he’s a hardware engineer to the core. Cook moved over to become executive chairman.

The outside world’s first reaction was basically unanimous. The biggest question marks around Apple right now are AI and software. Siri was rebuilt, delayed for more than two years and still not delivered, yet the board chose a hardware person as the next leader. If they wanted to make up the software shortfall, they brought in the person most skilled at tightening screws.

If you go through Apple’s own books, though, the board’s choice is actually quite coherent.

In the last quarter, Apple’s revenue was up 16% year over year, setting a record for the June quarter, and iPhone revenue rose 22%. The idea that products can’t sell isn’t something Apple has to worry about—at least for now. The pressure is in gross margin. After excluding tariff refund adjustments, Apple’s gross margin was 49.3% in the March quarter, fell to 48.1% in the June quarter, and guidance for the September quarter came in at 47% to 48%.

At the late-July earnings call, CFO Parekh laid out the explanation plainly: changes in memory costs can account for more than 100% of the quarter-over-quarter decline in gross margin. Other cost items added together are, in a way, helping—the entire drop is driven by memory, and it even dragged a bit further.

In that same call, Cook described memory price hikes as a once-in-a-century flood. As a result, Apple reluctantly raised the prices of the iPad and Mac. He added one more, more important point: looking beyond September, memory market pricing would continue to rise, and the impact on the business could be even bigger.

The reason for shortages tells the story even better. Cook said it was a demand-forecasting issue. iPhone and Mac are selling far better than the company’s own expectations. Apple could get the supply—what it didn’t do was order enough at the start. Meanwhile, memory manufacturers are prioritizing AI data centers, where profit margins are higher.

After taking over, after the swap, Apple will be fighting a battle for materials and capacity over the next two years. He has to bring negotiations with Samsung, SK hynix, and Micron to squeeze down prices and secure allocations. Internally, he also needs to cut redundant single-device memory usage during the design stage and push Apple’s in-house chips forward to offset the external price increases. All of these are hardware-engineering tasks. It’s far more fitting to let someone who has spent more than twenty years doing hardware at Apple manage this than to put a model expert in charge.

As for AI, Apple has already answered with actions. The rebuilt Siri’s underlying layer uses Google’s Gemini, costing roughly $1 billion per year. In that model battle, Apple doesn’t plan to win purely by doing it all itself—by buying this layer, it keeps its strength focused on devices, chips, and the privacy architecture. Tenas’s appointment aligns with this strategy.

Wall Street has meaningful disagreement with this logic. Wedbush’s Dan Ives raised his 12-month target price to $400 before the new CEO took office—on the optimistic end among mainstream institutions. His bet is on the iPhone 18 cycle, plus the re-acceleration in services revenue once Siri is rolled out. Jefferies’s Edison Lee went in the opposite direction: in August he cut Apple from “hold” to “underperform,” and cut the target price to $263.66. His basis is supply-chain research: problems hit the yield rate for a fully glass iPhone slated for 2027, and both memory costs and the slower pace of AI deployment are lagging as well.

My ranking differs somewhat from both sides.

If you trade this leadership change as an inflection point in the AI narrative, the order flips. Apple’s toughest constraint right now is cost. The layer of model capability is already paid for. Whether it’s done well or not will be answered by the September 9 keynote, but whatever it reflects in the financial statements won’t show up until next year. Memory is the kind of item that’s actively eating into gross margin right now—and, by Cook’s own words, it’s still worsening.

Apple also did another underappreciated thing. At the spring earnings call, Parekh announced that the company was abandoning its long-standing goal of net cash neutrality and switching to independently assessing cash and debt. That move loosens the leash on the balance sheet—going forward, it means more flexibility to carry out large AI acquisitions or ramp up R&D spending, with one less layer of accounting constraint. Uncapping the ammunition at the moment the torch is passed to the new CEO doesn’t feel coincidental.

The counterargument holds up as well. On September 1, Apple strengthened on its own even as the broader market fell. That was due to the interest-rate logic—there was essentially no fundamental change that day. When the yield on 10-year U.S. Treasuries pushes higher, capital tends to seek companies that don’t need to expand by taking on new debt. Apple doesn’t have a capital expenditure cycle like that, and its free cash flow is thick—so in this environment it’s naturally a safe haven. This logic can temporarily outweigh the gross margin issue, but it doesn’t solve gross margin.

Valuation is also indeed not cheap. Based on the data cited by 24/7 Wall St. on September 1, the Street consensus target price was still below the then-current share price. In other words, the good news had largely already been priced in. Analyst ratings are nearly unanimous on the buy side—this kind of consensus itself signals that expectations aren’t low. The sentiment around this #美股 area is closing the distance with the fundamentals.

What could overturn this view? If after September 9, Apple’s December-quarter gross margin guidance returns to above 49%, it would suggest I overestimated the memory constraint. If after the new Siri is delivered, services revenue can truly re-accelerate, then AI would be the main storyline for Apple next—and I would need to overturn the ranking I’m making today.

$AAPLB is now quoting at $326.71, already above the leadership-swap headline.

Over the next two months, rather than fixating on what the new CEO says, you may want to pay more attention to the tone Apple sets for the new Siri at the September 9 keynote, and where the December-quarter gross margin guidance lands in the late-October earnings report. That later number is more truthful than the keynote.

After SGP-0002 passed, I went to a Solana mainnet RPC node to check the current inflation parameters. The “decay rate” field still returned the same value as before. The winning side has already celebrated on Twitter, and not a single on-chain rule has been changed.

The vote approved was an authorization—nothing in the protocol itself had moved yet. This interim period determines when the issuance of $SOL will truly start decreasing.

Since genesis, SOL’s issuance follows a downward curve: each year it decays by 15% from the previous year, until it hits a floor of 1.5%. SGP-0002 doubles the decay speed to 30%, leaves the floor unchanged, and cuts the time to reach the endpoint in half. Projecting from the mainnet’s current inflation rate, the old rule would take six more years to reach the floor, while the new rule reaches it in three. For the portion that is reduced, the model in the proposal estimates about 18.90 million fewer SOL issued over the next six years. That figure comes from two Helius engineers; since nothing has been delivered on-chain yet, the quoted number needs to include that caveat.

The vote’s drama is bigger than the proposal itself. The “yes” side pushed through the supermajority threshold by locking in more than two-thirds, with only a 0.334 percentage-point margin. More than an hour before the close, the “yes” votes were still far behind. The batch of delegated votes that Kraken had on hand first flipped from support to opposition, pushing the vote counts below the line; at the last moment, it moved the vast majority back into the “yes” column. Helius CEO Mert Mumtaz said that in the final hours he had been calling nonstop to pull votes, and the votes came in during the last few seconds.

Setting a two-thirds threshold on a network where stake is concentrated effectively hands the outcome to real-time judgment by seven or eight institutions. This time it tilted toward a direction of production reduction—next time, it may not.

Figment and Everstake, two major staking service providers, voted against. Everstake has publicly stated the reasoning: the pace of changes is too fast, small validators will be squeezed out disproportionately, and both staking participation and delegators will face pressure. That argument makes sense from the validator’s position—when returns are compressed, the first to be pushed out are indeed the small, cost-inflexible nodes.

Putting the conflict of interest on full display was Solana Company. This listed company publicly announced its opposition to SGP-0002 on August 21, and in its second-quarter revenue, 99.4% came from staking rewards on its own holdings. It also holds a large amount of delegated stake for vote-by-proxy; its stance on the production-reduction vote needs no explanation. Governance rules allow this. It’s not exactly “cheating,” but people who delegate their votes have the right to know what ledger backs that vote.

During the voting period, Mert’s claim was that certain so-called “stakeholders” make money for themselves by diluting token holders through additional issuance—his model is said not to hold up. The proposal document even describes the algorithm used by his camp. On the whole network, 41% of validators charge a 0% commission on the issuance portion; reducing issuance won’t hit this group of nodes. The remaining nodes have their decay rate adjusted—since the current-period interest rate doesn’t change, revenue won’t collapse overnight.

What I don’t buy is reading this vote as a direct deflationary positive for $SOL . Spreading 18.90 million coins across six years, placing them into SOL’s daily trading float is just a drop in the bucket. These three days of market action didn’t give it face either: SOL kept giving back gains, sliding to 101.74, down 3.17% over 24 hours—tracking the broader-market rhythm.

The value of this vote is elsewhere. In the same batch of proposals, a constitution-related proposal passed with a high vote count, while SGP-0003—which changed transaction fees into a resource-pricing model and significantly increased the amount burned—only got 53.9% and failed. Taken together, the community’s message is quite concrete: it’s willing to mint fewer coins, but unwilling to turn SOL from a cheap execution chain into an asset that captures value by burning fees. Reducing issuance is essentially re-cutting the cake between token holders and validators; changing fees would shift costs onto application operators, and that faces much stronger resistance. This signal is more useful than the 18.90 million coins for judging Solana’s value-capture path.

Back to the parameter at the start that hasn’t changed yet. The protocol needs to use the technical proposal SIMD-0550 to make the change. In the warehouse’s status fields for this document, as of today it still says “Review.” The field that records the enable/disable switch value is still a placeholder waiting to be filled.

On Anza’s side, in early August they already merged the implementation into the Agave mainline. Their approach is to add a feature switch, and to re-anchor the curve at the epoch boundary when it becomes effective.

The problem is dated August 24. Two engineers from Jump and Firedancer opened an issue in the proposal repository. SIMD-0550 requires all clients to compute the results of exponentiation bit-for-bit consistently. But under IEEE 754, this operation is implementation-defined. Swapping the same code to a different CPU architecture, a different libc, or a different compiler version can lead to mismatched mantissas. The wording from the Firedancer engineer is that the feature switch should be blocked, and the suggestion is to revise the proposal so it uses no floating point at all.

In a single-client era, this wouldn’t be a big deal. In the current multi-client parallel stage, it becomes a consensus safety problem. Issuance rewards would go into the bank hash; if two clients compute a difference of even one bit, the chain forks. Anza’s engineer later proposed a patch that removes floating point from the inflation and rent paths. So far, it has only received a basically approving response that said to wait for others to review; it hasn’t been merged.

The governance layer approved it, but the spec layer is still waiting to change things. The client implementation is being questioned as not safe to launch. So activation can’t really be discussed yet. Until this path is completed, not a single coin of issuance will be reduced. #Solana

The circumstances that could overturn this assessment are also easy to list. Once the patch is merged, a new version is released, and validators are activated—if the staking ratio and the number of validators haven’t fallen—then the opponents’ concerns would be overblown, and I’d scale down the weight I gave to their worries. Conversely, if the switch gets stuck for months, or after activation small nodes exit in batches, then this close call victory would turn into a headache.

If you want to personally monitor this line, you can check Solana mainnet RPC’s getInflationGovernor interface. The taper field in there is still the old value for now. Only when it doubles will the production reduction truly begin.

Coinbase’s official account posted a weekend image with only a few words in its promise: no selling the coins, no additional margin calls, and no taxable event. The day before, Brian Armstrong’s post was even more direct: get the house, and keep your exposure to Bitcoin. The product officially opened last week for qualifying U.S. homebuyers. I checked every one of these lines against the original wording; they all hold up. It’s just that they “stand” in a way that differs from what most rewrites in the Chinese-speaking market say.

What gets most crushed here is the structure. There are two loans. The first follows Fannie/Freddie’s standard compliance-approved framing: the collateral is the house, and it has nothing to do with crypto assets. The second is a separate down-payment loan, used to cover the cash down payment that the buyer can’t come up with; the collateral is the Bitcoin pledged for it, and it also adds a second mortgage lien on the house. Hardly anyone mentions that second lien, but it determines the order of default: Better can go after the house first; the coins are merely a later-layer backstop. The interest rates and amortization term are the same for both, and each month they’re merged into a single repayment.

By putting all Bitcoin volatility into the second loan—carried by Better—the first loan remains the kind of compliant asset that Fannie/Freddie can simply take and hold. Because of this design, the product doesn’t need to wait for any new regulatory guidance to be finalized before it can operate. This layer is especially easy to mix up. FHFA chair Pulte signed a directive at the end of June last year instructing the two housing agencies to treat crypto assets held on regulated U.S. exchanges as part of the reserves for mortgage qualification, without needing to convert them into U.S. dollars first. Even up to mid-year this year, that line still has no final guidance. Including Senators Warren and Sanders, several lawmakers sent letters questioning whether it would shake up the housing market. Whether Fannie/Freddie recognizes the coins, and whether Better uses the coins for a down-payment pledge loan, are two separate matters that don’t depend on each other. Writing that Fannie/Freddie started accepting Bitcoin directly is wrong.

The “no selling the coins” line holds—because the price is embedded in the loan-to-value ratio. The down-payment loan requires the pledged Bitcoin to be no less than 250% of the loan amount. If you want to borrow $100,000 for the down payment, you must pledge $250,000 worth of Bitcoin. The coins are transferred into Better’s custody account via Coinbase Prime and locked until the loan is repaid or refinanced. At the current price of $78,075 for $BTC , a $100,000 down-payment loan would consume more than three whole Bitcoins. Those more-than-three Bitcoins may be unable to move for as long as the next thirty years, and they can’t be used elsewhere as collateral. It doesn’t generate interest, but the years they’re locked up have a cost—just not one that shows up in the marketing materials.

The “no additional margin calls” line is even more counterintuitive. A drop in the coin price by itself doesn’t change the mortgage terms, and it won’t trigger a forced liquidation because the collateral value shrinks. The trigger for liquidation is switched to repayment delinquency: after being late for sixty days, Better would then be entitled to dispose of the pledged Bitcoin.

In this way, the risk exposure shifts from price to cash flow. It doesn’t care how much your coins fall; it only cares whether you can make your monthly payment each month. The 250% is what you paid for that promise—an upfront over-collateralization replacing an entire dynamic margining/mark-to-market backstop mechanism.

For borrowers, that’s a good thing, but the risk just changes shape. The worst path is when a major drop in coin price happens at the same time as your income dries up: two months of delinquency, Better sells your coins at the bottom, and that second lien on the house remains. People heavily concentrated in holding coins especially can’t avoid this combination; their income source often ties to the same industry. When things are good, both sides have breathing room; when trouble hits, they hit together. The “no additional margin calls” part is true. The missing half is that the liquidation switch has been moved onto your paycheck.

The “no taxable event” line also fails under the same path. The pledge itself indeed doesn’t constitute a disposal. But at the moment of a passive closeout, it becomes a real sale; the capital gains tax you owe is not one cent less, and it happens when your cash is tight—precisely when you least want to sell.

In Better’s disclosed numbers, one figure is even more telling than the 250%: among its pre-approved customers, 41% have both income and credit that pass the bar, and what blocks them isn’t affordability—it’s simply that they can’t produce the down-payment cash. Earn enough and you can borrow; the money is all in the coins. The intended loan size corresponding to the waitlist stage exceeds $260 million. These are the company’s own disclosed figures. They may work as marketing numbers, but the 41% points to a real gap.

For a few institutions that do non-compliant loans, their attitudes are close to one another: they all think adding crypto assets to mortgage qualification is inevitable, and they all stress that their version is conservative. Rate’s Kate Amor directly calls their product a conservative non-compliant one. The objections concentrate on the two-agency (Fannie/Freddie) line—not on this product. A Federal Reserve Bank of New York study in 2024 mentioned that when crypto assets are under pressure in traditional markets, their volatility is comparable to traditional assets; once the crypto market starts breaking on its own, the volatility will be much more severe.

Structurally, this product is conservative to the point that it doesn’t really look like a typical crypto product. The first source of repayment is the house; the second is your income. Bitcoin comes third—and you’re required to provide thickness of 2.5 times. The criticism that it brings crypto volatility into the housing market misses the mark: on the compliant mortgage side, it doesn’t touch the coins at all. What should be watched is the other side: it wraps a 30-year liability that needs cash-flow discipline into an option where coin holders supposedly don’t have to make trade-offs.

This judgment breaks down in two situations. First, as the loan-to-value ratio is competitively squeezed as peers enter the market—say 250% drops to one hundred-something—the thickness provided by over-collateralization may no longer hold. Second, if these down-payment loans begin getting packaged and sold off rather than staying on Better’s own books, once the risk is transferred away, the incentive to maintain strict issuance standards also moves away. Both of these can be inferred from public product terms and securitization trends; you don’t need to wait for something to happen.

This round around $BTC started on August 19 and, over ten days, climbed from just over $64,000 to $81,479 intraday on August 28; it has now come back to $78,075. At this level, a 250% collateral requirement looks comfortable; three weeks earlier, at that earlier price level, it looked like an entirely different story. The natural peak in signing such products falls when the collateral looks most plentiful—and that’s also when it’s most expensive.

To judge whether the numbers add up, you can start with a crude calculation: value the portion of coins you must pledge assuming they can’t move for thirty years, then ask yourself: if in any year in between you need to use it, do you have other money on hand that can cover the need? The answer is far more important than how high Bitcoin might rise.

Amazon’s net profit last quarter jumped sharply—so sharply it looks like AWS suddenly switched to a money-printing machine. But you can’t read this number that way. Most of it comes from a mark-to-market revaluation of Anthropic equity the company already holds, not from selling cloud services.

Strip that out: in Q2, operating profit was $27.5 billion—that’s the money actually earned by the business itself. This matters because Amazon has signed another bill this week that will be gradually digested through operating profit.

On August 26, AWS and Nvidia announced an additional 2 million GPUs, with delivery scheduled for 2027 to 2028. Add that to the 1 million GPUs earlier this year, and the committed total is raised to over 3 million. The models cover Blackwell Ultra, Rubin, and Rubin Ultra. After the U.S. stock market opened on Friday, the market’s reaction to this order was: buyers up, suppliers down. The reading at $AMZNB at 20:18 is $266.21, up 3.44% over 24 hours. That rally is concentrated in the few hours of U.S. trading on Friday. Since Saturday, the order book has barely moved, with only a few hundred thousand dollars in volume—so it can’t serve as a reliable “sentiment” read. In the same window, $NVDAB is -3.54%.

Nvidia’s red candle has little to do with this order. Late on August 27, The Wall Street Journal reported that Nvidia paused some transactions in a financing program it launched in July. The program previously disclosed a commitment size of about $36 billion. The mechanism is: Nvidia first extends credit to small- and mid-sized AI cloud companies so they can afford the chips. Once those cloud companies rent out compute, the revenue portion above the agreed cost is shared back with Nvidia. If they fail to rent out capacity, the capacity is taken back by Nvidia. The report cites internal antitrust concerns and partners’ dissatisfaction with restrictions on customers. Externally, Nvidia still says the program is running. Some media have questioned the report—this is the point to record for now.

Two developments landing in the same week split people buying the cards into two groups. Amazon’s 3 million GPUs are funded from the company’s own cash flow. The batch for smaller cloud firms depends on the supplier first standing behind them so the purchase can go through. The divergence on Friday comes down to this difference.

So let’s go back to Amazon itself. It can afford to buy—no doubt. The harder question is how cheaply (or expensively) the items it buys can be valued.

In Q2, AWS revenue was $42.2 billion, up 37% year over year. Operating profit margin was 39.4%, versus 32.9% in the same period last year. When money is being spent most aggressively, the profit margin actually widened by more than six percentage points—exactly the opposite of what intuition suggests.

The reason is the speed gap between two lines. In the same quarter, Amazon’s depreciation and amortization was $20 billion, versus $15.2 billion in the prior-year quarter—a 30% increase. Depreciation is indeed rising, but AWS revenue is rising even faster. What people call the “depreciation wall” is when the timing comes for the period when revenue growth slows below depreciation growth. Those two lines have not crossed yet. #AI infrastructure

The cash-side tightness is real. Rolling 12-month operating cash flow was $161.4 billion. Free cash flow was -$7.6 billion; last year, the figure was positive. The gap comes from spending on land, building data centers, and buying chips—funding the company says in its earnings materials is mainly aimed at artificial intelligence. Management raised its 2026 capital expenditure guidance to $220 billion, with the direct rationale being rising memory prices.

There is one detail that better shows how management thinks than the guidance number itself. Starting January 1, 2025, Amazon changed the depreciation life of some servers and network equipment from six years back to five. The reason given was that AI makes technology cycles move faster. The debate outside about whether compute-asset lifespans are overestimated has been going on for more than a year. Before the dispute grew too big, Amazon tightened its own assumptions. This move is a negative for current-period earnings, and the willingness to do it suggests the company isn’t very optimistic about how long its equipment will actually last.

The external disagreement is exactly concentrated on that assumption. On August 28, Evercore ISI raised its price target for Amazon while maintaining its bullish view. Around the same time, Rosenblatt Securities initiated coverage with a Buy rating. Both are betting that AWS growth hasn’t reached its end yet. The strongest argument from the bearish camp comes from Michael Burry, who has long argued that the real usable life of compute assets is shorter than the accounting assumptions by a noticeable margin. If that’s true, the depreciation expense of these mega-scale players over the past few years has been systematically underestimated, meaning part of the book profit is essentially “borrowed.”

I lean toward the bullish side. The reasons differ from the seller’s, though. Amazon’s ability to pay isn’t the issue: $161.4 billion in operating cash flow plus room for financing easily covers the 3 million chips. The risk falls on the speed gap between revenue growth and depreciation. Depreciation is a cost locked in once contracts are signed; it hits steadily each quarter. AWS revenue growth depends on whether customer budgets keep up. In this 37% figure, how much comes from long-term commitments versus phase-based surges like AI training—Amazon hasn’t broken it down. That’s the part I’m least certain about right now.

The only read that could make me change my mind is whether AWS can defend operating profit margin around 39%. If it holds, it suggests the additional depreciation is still being absorbed by revenue. If it drops, it indicates this round of capital spending is starting to eat into profit, and the whole algorithm behind these assumptions has to be rebuilt.

Looking on the optimistic side, the reasons aren’t weak either. If AI demand in 2027 to 2028 really is as Nvidia management described—long-term supply shortfalls—then every additional chip bought now will become revenue earlier, and depreciation pressure will be diluted. In that case, worrying about margins now is just overthinking. Amazon itself has also said that by 2027, capacity won’t be able to catch up with demand. If that holds, then the earlier concerns are indeed excessive.

$AMZNB hasn’t had much pricing information over the past two days. Waiting until after the U.S. stock market opens on Monday will make the traded volume meaningful. If you want to follow this line, you can look at the two lines in the next quarterly report for AWS—operating profit margin and depreciation/amortization—which are more useful than obsessing over day-to-day price moves.

When it comes to making Bitcoin quantum-resistant, several different paths suddenly emerged this week, and they are completely different in how they’re carried out. The first one to go viral is also the easiest to understand: Bitcoin’s rules don’t change at all—repeatedly re-compute the signature for a payment until a quantum computer can’t do anything about it, and then hand the transaction straight to the miners for packaging.

The premise is a bit awkward: ordinary nodes simply don’t relay transactions in this format—you have to knock on a mining pool’s door yourself.

This transaction landed in block number 964,199. The scheme was designed by Avihu Levy of StarkWare, and his colleague Tomer Giladi routed it through MARA’s Slipstream channel. I pulled it apart on-chain and checked it: the input protected by the quantum-resistance mechanism is 10,000 sats, which at today’s price comes to less than eight dollars. Levy has explained the principle: the wallet doesn’t accept the first valid signature it computes. Instead, it keeps generating candidate signatures over and over until it finds the one with the right shape—this process consumes several hours of computing power.

Eight dollars, several hours. Pulling this comparison out isn’t meant to mock; it actually shows what this proof demonstrated. Under Bitcoin’s consensus rules today, a spend that does not leak a public key throughout can indeed be packed into the main net.

The scheme’s boundaries are something StarkWare itself describes more honestly than the people who circulated it. The company states clearly that QSB did not make Bitcoin quantum-resistant. It protects only certain transactions. Specifically, it only works for addresses whose public keys have already been exposed on-chain; it can’t save others. CEO Eli Ben-Sasson puts it even more plainly: he still wants Bitcoin to do a soft fork, and he believes it will happen in the end.

The trouble is right here. Bitcoin’s quantum risk isn’t distributed evenly across all coins—it is concentrated only on the subset of addresses whose public keys are already written into blocks. An address that has spent money belongs to this category. As of March 1 of this year, more than one-third of all Bitcoin on the network has already exposed public keys on-chain. These coins won’t become automatically safe just because there’s a new way to spend them. Either the owners actively move them, or they just sit there indefinitely waiting. QSB is a tool prepared for the portion that hasn’t exposed public keys; that’s exactly the group that most needs rescue, but it can’t reach them.

The second path goes after this gap. In the same week, Blockstream’s Jonas Nick formally published BIP for SHRINCS. This is the first quantum-resistant signature scheme specifically tailored by trimming according to Bitcoin’s ledger structure. The foundation is still SHA-256, and it doesn’t rely on any new mathematical assumptions. The cost is made explicit: today a Schnorr signature is 64 bytes; SHRINCS has a minimum of 548 bytes, and in the worst case it can grow to 4,619 bytes. It also requires state: with the same private key, each time you sign the key grows a bit longer. If the device is lost, you need a fallback transaction of more than 5,000 bytes to recover the funds. In the BIP document, there’s also a line that has not been deleted: the security proof hasn’t been completed.

The third path goes even further. In the same week, Blockstream also released an evaluation of lattice-based signatures. Falcon-1024 is the most space-efficient in that category: public key plus signature together totals 3,073 bytes. But the research team itself didn’t recommend deploying it right now, and NIST’s standard text isn’t finalized yet. Their suggested order is: use the hash-based approach first, and only consider a hybrid once the Falcon standard is settled.

Once the routes are laid out, the shared point can’t be hidden. Besides QSB, the other two paths require changing the consensus layer. QSB doesn’t because it bypasses the entire P2P network: nodes don’t recognize such transactions, so miners have to receive and package them separately. Engineering-wise, Bitcoin today isn’t short of answers—it’s short of someone who has the authority to decide for the one-third of coins.

This contradiction has already been brought to the surface this year. In February, BIP-360 was merged into the official repository, defining Bitcoin’s first quantum-resistant address type. In April, Jameson Lopp and five other developers published BIP-361, setting a five-year sunset period for old signature types. Coins that haven’t been moved by the deadline will no longer be recognized by the network as spendable, including the batch widely believed to be Satoshi’s. Adam Back is explicitly against forced freezing; he argues that quantum-resistant functionality should be made an optional feature now, so people can move their own coins. The most accurate summary came from Marin Ivezic, who works on post-quantum security; he said the true constraint for Bitcoin’s quantum migration isn’t cryptography—it’s governance.

I agree with that judgment, and this week’s news provides a perfect footnote. Cryptographers have finished the multiple-choice part: QSB is what can already be used; SHRINCS is what can be brought into the protocol; and Falcon is the option that saves space. The remaining controversy no longer belongs to technical selection—it’s whether to set a deadline for the coins owned by some people. Bitcoin’s governance structure is capable of adding features—Taproot is proof of that. But when it comes to taking away rights, it has never succeeded. The original design was meant to block exactly this kind of thing.

This judgment can be falsified. Over the next few months, if BIP-360 or SHRINCS enters substantial activation discussions and rejects the kind of signaling schedule Taproot used that year, the governance bottleneck might not be as stuck as I imagine. Another signal could be even more direct on-chain: if large addresses that haven’t moved in more than ten years—and whose public keys are already exposed—start relocating in bulk, then the debate about freezing versus not freezing will automatically be downgraded. Neither of these has happened yet.

And there’s no need to be scared by this week alone. The market’s starting point for quantum anxiety is late March and June. Google Quantum AI improved the resource estimation for Shor’s algorithm on elliptic curves by an order of magnitude, and Justin Drake’s long write-up spread it widely in the community. Resource estimation improvement is not the same as actually building machines—the former only shifts the timetable forward a bit. #Bitcoin is now 79,891, and overall this week it’s still moving upward, basically unrelated to the quantum timeline. The value of $BTC is still running along with macro factors and the ETF schedule.

If you want to do something for yourself these days, you can check whether the commonly used address you control has spent funds on-chain. If it has, that means the public key is already exposed, and later—no matter which path Bitcoin chooses—the addresses that will need to be moved proactively are exactly this kind. It’s still far from that day, but knowing which side you’re on is more useful than remembering which week someone proposed which scheme.

The hardest-to-read part of Nvidia’s quarterly report isn’t in the income statement. The fact that revenue has doubled is already baked in by the time the seller-models start their calls—there’s no real suspense. The disagreement is concentrated in the footnotes on the following pages: whom the company has guaranteed leases for, how many years’ worth of purchase orders it has signed, and which customers it has extended payment terms to. Put these three items together, and they explain how the money from this round of AI capex circulates better than any year-over-year figure ever could.

Let’s get what’s on the surface out of the way. In the second quarter, revenue was $96.2 billion, with data center accounting for more than 90%. Gross margin held steady at above 70%, and the third-quarter guidance came in at $108 billion, higher than the Street’s consensus. These numbers aren’t controversial—they just confirm something that was already known: Blackwell Ultra is still ramping up, and whatever capacity it can produce, it sells.

The new element was a table CFO Colette Kress chose to present proactively. Nvidia’s supply and production commitments jumped from $119 billion in the prior quarter to $279 billion, an increase of $160 billion in a single quarter. Her explanation was that the money is mainly going into memory procurement. The explanation is plausible, but the implications are heavier than what it sounds like. Memory contracts are long-term—when the deal is signed, both price and quantity are locked. In effect, Nvidia is betting on the shape of demand over the next three years, and it has already paid a deposit. On the call, Kress said that instead of letting this table become a hanging question, it should be made clear outright.

Accounts receivable at quarter-end was $63.1 billion, and DSO stretched from 45 days the previous quarter to 60 days. In the 10-Q, the company spells out the standard: for large purchases from investment-grade customers, the payment terms can be extended to 90 days, and up to a year, to align with customers’ large data center construction. The same filing also includes another line noting that the combined receivables balance for five direct customers accounts for 70%.

Read these two lines together, and the picture comes into focus. Chips ship, revenue is recognized in the period, but the cash arrives only three months to a year later—and the unpaid amounts are concentrated among a few customers. Nvidia is using its own balance sheet to help customers finance working capital turnover. This isn’t automatically “bad debt,” and investment-grade customers will likely pay—but it shifts part of those customers’ credit risk onto Nvidia’s books.

In August, Nvidia also signed a guarantee with a cap of $105 billion. The guarantee covers an SB Energy campus in Pike County, Ohio, supporting the credit for land, power, and the factory, with the tenant being an OpenAI affiliate entity. The scale is roughly 4.25 gigawatts. This guarantee isn’t an investment, and it isn’t a cash outlay. It will only take effect gradually as the data center is built in phases and the lease becomes effective. The first tranche is expected in fiscal 2029; each time OpenAI pays rent for a tranche, the exposure decreases a bit. In exchange, the campus will deploy only Nvidia equipment.

Huang Renxun has long denied that this is cyclical financing. His rationale is that rent is paid by OpenAI, and Nvidia locks in resources only where it can see demand. On the call, Kress added one more point: the demand such cooperation can generate is roughly a quarter of next year’s business, and Nvidia’s platform is general-purpose and durable, making the risk therefore controllable.

The counterargument also comes with names and titles. Bill Birmingham of Rex Financial said the guarantee amount shrank from the more than $200 billion level previously rumored in the market down to $105 billion. The market read it as demand shrinking, not as risk falling. Nvidia, he said, lost $250 billion in market value because of this. Melissa Otto of S&P Global Visible Alpha took the other side; she said the whole market was shocked by the 70% figure.

The 70% refers to the annual guidance Huang Renxun gave as an exception during the call: revenue growth of 70% in fiscal 2028, while the Street consensus was only 44%. He also added that this was calculated based on supply capability, and that actual demand would be higher than that. Those words were the turning point in the trading that night. In the first hour after the earnings release, the spot order on Binance priced at $NVDAB was briefly smashed to just above 204. After the call started, it kept being pulled back, and within 24 hours it was up 3.2%. In the same #Nvidia earnings report, the income statement makes people tense, the forward-looking guidance makes people feel reassured, and what lies in between is how to read these footnotes.

My own judgment leans toward acknowledging Huang’s supply-logic explanation. Memory is the toughest bottleneck right now. Locking supply three years in advance makes sense commercially, and the $279 billion figure looks more like抢产能—grabbing capacity—than hard-building demand. But I don’t accept the claim that risk hasn’t changed. In the same quarter, Nvidia’s cash flow from operating activities was $24.1 billion, and GAAP net profit was $59.7 billion—more than double the gap. The shortfall mainly came from accounts receivable and inventory. It also issued $25 billion of senior unsecured bonds. For a company with ample cash, extending payment terms to customers while still needing to issue debt indicates that the funding pressure from this expansion is already starting to transfer onto the company itself.

There’s also an easy-to-miss accounting perspective. This quarter, GAAP earnings per share were $2.46, while non-GAAP was only $2.22. GAAP was actually higher. The difference came from $7.8 billion in equity investment gains, which were unrealized gains from Nvidia’s holdings of equity in AI companies. Non-GAAP excludes those. When the valuations of the companies Nvidia invests in rise, it directly lifts Nvidia’s reported book profit. This link is a plus item in an up-cycle—and when the direction reverses, it’s just as responsive.

Under what circumstances would I admit I’m wrong? If over the next one or two quarters DSO shrinks back toward the 45-days range, and operating cash flow catches up to net profit again, then this payment-term loosening would just be a timing difference for a few large orders. In that case, my concern would be overinterpretation. Conversely, if DSO keeps moving upward, and the concentration of receivables continues to exceed 70%, then the “investment-grade customers” wording will carry too much weight, and the market will eventually demand that Nvidia clearly disclose the names of these few customers.

The next quarter’s focus won’t be whether revenue can reach $108 billion—it will most likely. Watch how these tables in the 10-Q change: whether supply commitments add more, which direction DSO moves, and whether another name shows up in the guarantee schedule. The main text of the earnings report is written for everyone; the footnotes are written for people willing to spend an extra twenty minutes reading.

This year’s harshest research report on Circle on Wall Street—when it was published, it happened to coincide with the stock’s intra-year low. Morgan Stanley cut its rating to “Sell” (reduce holdings), slashed the price target by more than half; on the same day TD Cowen flipped the script and initiated a “Buy.” Three weeks later, the price of $CRCLB had left both firms’ targets in the dust. Needless to say, the bearish note was wrong; but even the bullish one didn’t keep up.

On August 3, Morgan Stanley’s James Faucette downgraded CRCL from Hold/Watch to Sell (reduce holdings), setting a target price of $38. On the same day, TD Cowen initiated coverage with a Buy rating and a target price of $82. That day CRCL closed at $60.35—the lowest day in three months. On Monday, on the NYSE, it closed at $87.72. On Binance’s spot market, $CRCLB is now quoted at $85.9, and in pre-market trading it has pulled back slightly, tracking the crypto-concept stocks.

This rally has little to do with reserve income. In Q2, Circle’s reserve income year over year increased by only a low double-digit percentage—or rather, just a single-digit gain. Its reserve yield was also lower than last year, and total revenue failed to beat market expectations. What the market bought were two other things. At the end of July, Circle fully acquired IBM’s blockchain patent portfolio, instantly becoming the U.S. company with the most blockchain patents in hand. Then on August 19 it also announced that Arc mainnet is scheduled for September 16. The day that news broke, CRCL’s trading volume more than doubled versus the prior day.

The money-making machine at Circle hasn’t been fixed yet. Today, USDC circulating supply stands at $73.8 billion—almost unchanged since the end of June. This year’s high was $79.6 billion on March 18. Coincidentally, CRCL’s own closing high this year also occurred on March 18, at $132.84. Both curves peaked on the same day. Before that, the market’s valuation of Circle was basically calculated based on circulating supply. Faucette’s downgrade landed right on that point. He cut his assumptions for USDC in 2027 and 2028 by 30% to 40%, and added a more unkind line: stablecoin usage is always concentrated in crypto trading and has not expanded into payments. Tokenized money market funds and tokenized deposits—both of these “things” steal both balances and fee rates.

Stopping circulating supply is only the surface. Circle’s trouble is buried in the cost line. In the same quarter, reserve income was $668 million, while distribution, trading, and other costs were $412 million. More than 60% of the money changes hands and gets paid out. The main recipient is Coinbase: all reserve income attributable to the portion of USDC held on the Coinbase platform goes entirely to Coinbase, while the two other counterparties split it fifty-fifty for USDC held elsewhere. At the end of June, about 30% of USDC was sitting on Coinbase’s books. This profit-sharing agreement was renewed in August under the original terms through 2029—effectively “welding shut” the cost structure for the next three years in advance.

Coinbase signed that agreement and is also one of the initiators of Open USD. This alliance includes more than 100 institutions—Visa, Mastercard, BlackRock, Stripe, Google are all in it. The rule of the game is: reserve earnings minus management fees are returned to the participating merchants. The slice of income that Circle and Tether have survived on by earning the spread has been carved out in this design. BlackRock is also a collaboration partner for Arc. These institutions sign on both sides; what they’re betting on is the #稳定币 track itself. As for who the issuer ends up being, they don’t care as much.

This year, the interest-rate line has actually been a tailwind. Warsh is scheduled to speak at Jackson Hole this Friday—his first public appearance since becoming chair of the Fed. Inflation hasn’t yet returned to target. In the July minutes, three votes favored tightening. Market pricing for a September rate cut has been drifting downward for weeks. Higher rates are good for Circle’s reserve income. If the shorts bet that rate cuts would thin the spread, then this year’s logic has not played out. The pressure comes from stalled circulating supply and that profit-sharing table, not from the Fed.

The market has shifted the way it values Circle—from an interest-rate ledger to a network-usage ledger—and in terms of direction, I agree. Faucette’s $38 target treats Circle as pure interest-rate beta and cut it too aggressively. Things like licenses, patents, and a clearing chain that is about to go live—none of that belongs in a spread model. But at $85, the market is already buying September 16 as the realization date. Arc going live is only the start of work; even if the testnet is bustling, it still hasn’t reached the point of charging.

The doubled full-year other-income guidance in Q2 included roughly $180 million from staged confirmations of Arc token pre-sales—one-time revenue. If you strip that out, look at whether other income in Q3 and Q4 can stand on its own. If it can, then this round of gains has a basis for accounting. If you strip it out and it’s still only in the low tens of millions range, then what you’re buying with $85 is simply the story.

On September 16, the Arc mainnet launch coincided with this round of the FOMC. If you want to follow the path of #Circle , you can compare the USDC circulating supply before and after with the actual settlement volume on Arc itself—it’s more useful than staring at the order book.