The first major event in October is two bid sheets.
Anthropic in the U.S. plans to list at a valuation of $2 trillion, and the roadshow schedule points to the week of November 9.
China's DeepSeek's target valuation for its second-round financing is 500 billion yuan, and at the same time, Citic Securities should prepare for the STAR Market.
Meanwhile, the average premium paid by Wall Street when buying new shares has fallen from 24% at the end of June to less than 1%.
So they have to sell within the same cooling window.
This is not a coincidence; the two sets of accounts are grabbing the same batch of money.
First, let's talk about the one in the United States; the ledger is also longer.
Anthropic plans to hold its pre-listing investors’ day on October 14, aiming to get listed before Thanksgiving.
It plans to raise up to about $100 billion, targeting a valuation of around $2 trillion.
Based on the post-money valuation from its $65 billion funding round in May, it has more than doubled again in six months.
First, let’s see where the losses actually came from.
The prospectus reports revenue of $4.59 billion in 2025, up 12-fold year over year.
That same year, its GAAP net loss was $41.97 billion.
Operating loss: $8.06 billion.
The loss on this item was under 3 billion last year.
The 42-billion figure is the one most likely to make the headline.
But it can mislead you, and in a very typical way.
Of the 41.97 billion, about 34 billion was a non-cash expense caused by the higher valuation of convertible instruments.
This money was never paid out; it’s only a movement on the books.
The actual operating loss was $8.06 billion.
Compute and infrastructure spending totaled $7.33 billion, accounting for 58% of operating expenses that year.
That’s also about three times the spending on the same item the year before.
By the end of last year, it had $20.28 billion in cash and short-term investments.
There’s genuinely no shortage of cash in the short term; that’s not in dispute.
What you really can’t take back isn’t the loss.
It’s a different bill.
The prospectus lists about $518 billion in commitments for cloud, compute, and infrastructure, with terms of roughly ten years.
About 80% of it is non-cancellable.
In other words, you have to pay whether you use it or not.
The industry term for this kind of clause is “take-or-pay.”
That works out to about $414.4 billion that can’t be withdrawn.
Spread over ten years, that’s about $51.8 billion a year, seven times last year’s compute spending.
That’s 25 times its cash on hand.
So this bill can only be paid for with financing and revenue.
That’s the most expensive part of the entire document.
Media outlets interpret how this commitment is allocated differently.
Breaking down the disclosed cloud contracts, Google Cloud and Amazon alone account for $300 billion.
Another breakdown by payee shows that Broadcom has the largest equipment lease obligations, at about $161.2 billion.
Both readings are correct; they just use different accounting bases.
One counts cloud services; the other counts equipment leases.
So don’t treat it as spending for a single year.
Now, the China filing.
DeepSeek launched its first external funding round in April and closed it in June, raising about 50 billion yuan.
That round valued the company at more than 350 billion yuan post-money.
Founder Liang Wenfeng contributed about 20 billion yuan personally, while Tencent contributed about 10 billion yuan.
Companies in the CATL ecosystem contributed about 5 billion yuan.
After the corporate registration changes in July, the National AI Industry Investment Fund was added to the shareholder register.
It’s the only one among these investors with direct voting rights.
The target valuation for the second round is 500 billion yuan, with completion planned by the end of October at the latest.
In U.S. dollar terms, the target for this round is about $75 billion.
CITIC Securities has begun due diligence, but the two sides have yet to sign a sponsorship agreement.
The Shanghai Stock Exchange’s records also show no filing from it so far.
So it’s still a long way from ringing the opening bell.
It plans to file as early as the end of this year, with a listing targeted for 2027.
In June, the CSRC extended the STAR Market’s fifth listing standard to artificial intelligence.
So this path has only just opened up.
Its problem has never been a lack of money. It’s something else.
It’s that revenue can’t keep up with the pace of investment.
DeepSeek’s revenue in the first seven months of this year was about 475 million yuan, roughly 10 times its total for last year.
Net loss over the same period was about 715 million yuan.
Its API business has a gross margin of 82.9%, higher than Anthropic’s and OpenAI’s over the same period.
But AI infrastructure spending in the first seven months was about 1.1 billion yuan.
This expense was only about 1.2 billion yuan last year.
So spending surged sharply during that year.
What it wants now is market share, not margins.
In other words, it spent 11 billion yuan and brought in less than 500 million yuan in revenue.
At an investor meeting, Liang Wenfeng put its annualized revenue run rate at $1 billion.
On the same basis, that’s about 6.7 billion yuan.
Just a few months ago, this figure was under $500 million, so it really has doubled.
After the August price adjustment, customers’ usage costs rose to more than 2.3 times their previous level.
His view is that the price increase hasn’t reduced the number of customers.
In September, V4.1-Flash cut prices by about 60% again.
So it’s raising prices on one hand while cutting them on the other.
Both sides are protecting their scale. That’s what a cash-burning posture looks like.
It uses more than 70% of its compute for model training.
Less than 30% is used for inference.
Huawei could begin delivering training chips as early as the fourth quarter of this year.
Now we can put the two sets of figures side by side and compare them.
For the U.S. filing, $2 trillion divided by annualized revenue of $65 billion gives a multiple of 30.8x.
For the China filing, 500 billion yuan divided by an annualized revenue run rate of 6.7 billion yuan gives a multiple of 74.5x.
Based on annualized revenue, the asking valuation for the China filing is more than twice that of the U.S. one.
If you use recognized revenue instead, the multiple jumps to over 600x.
That shows why the basis of comparison matters more than the story.
For the same company, the multiple is 74x on an annualized basis and over 600x on audited figures.
The key is to compare again whether the compute spending is paying off.
Anthropic spent $7.33 billion on compute last year, 1.6 times its revenue of $4.59 billion.
DeepSeek spent 11 billion on infrastructure in the first seven months, against revenue of 475 million over the same period—that’s 23 times as much.
The two figures differ by more than an order of magnitude; they’re on completely different scales.
But you can’t draw a conclusion from this comparison alone.
Because the former is recorded as an operating expense.
The latter includes purchases of chips and servers.
The difference in scale still holds and can’t be ignored.
One is already covering its compute costs with revenue.
One is still using financing to cover its compute costs.
That’s the difference in scale I mentioned earlier.
However, compute has always been cheaper in China on a unit-cost basis.
According to third-party interviews, U.S. large language models cost more than ten times as much to run as their Chinese counterparts.
The two companies have made opposite choices in how they present their valuations.
Anthropic uses annualized revenue.
And it admits that nearly a quarter of last year’s revenue came from just two customers.
It also acknowledges that many major customers don’t have long-term contracts.
DeepSeek’s position is that it will file using a full year of financial statements.
Rather than anchoring its valuation to ARR.
If it really files using full-year financial statements, the valuation benchmark for AI stocks in China’s A-share market will shift back to auditable revenue.
Outside investors don’t get real control at either company.
Anthropic will continue to operate as a Delaware public benefit corporation.
The seven co-founders hold 50.1% of the voting rights through Founder LLC.
On DeepSeek’s side, most investors hold their stakes indirectly through Ningbo Cheng’en.
It’s a limited partnership with Liang Wenfeng as its general partner.
Investors have no direct voting rights and are also subject to a five-year lockup.
Intermediaries also charge upfront fees of 6% to 15%.
Some also demand a share of as much as 40% of the excess returns.
An investor at a Hong Kong family office said he had recently received DeepSeek allocations through eight different channels.
So the cost of getting in early isn’t just paying a little more.
It also adds a counterparty layer you can’t investigate at all.
That’s the hidden cost the secondary market can’t see.
Now let’s look at the funding side.
BlackRock fixed-income CIO Rick Rieder is selling stocks and buying bonds.
His reason: investment-grade bonds yielding 7% to 8% can already beat his expected stock returns.
He downgraded the stock to B−, but kept his ratings on chips and memory.
The 10-year Treasury yield was at 5.167% last week.
In September, the Federal Reserve raised interest rates to a range of 3.75% to 4%.
He calculated that every 100-basis-point rise in interest rates adds hundreds of billions of dollars in costs to the U.S. government.
Mortgage rates have already been pushed up to 7.45%.
That’s why he describes the U.S. housing market as frozen.
AI stocks valued at $2 trillion are competing with bonds yielding 8% for capital.
Others have already walked this mega-IPO path once before.
When SpaceX goes public in June, its valuation will be $1.77 trillion, with an offering price of $135 a share.
It gained 19% on its first day, closing at $160.
It then surged to a high of $225.64, before falling below its offering price for the first time in July.
On July 17, it closed at $123.99.
That was its first close below the offering price since it went public.
At its peak, short interest was close to 30% of the float.
As of mid-September settlement, 162 million shares were still short.
It closed at $158.96 this week, up 7.35% on the day.
Above the offering price, but below the peak.
This is what the first year of a mega-IPO usually looks like.
The conditions for making money on IPO subscriptions are disappearing at the same time.
The smart-ring company Oura has postponed its IPO.
It originally planned to issue 50 million shares, priced above $40.
At the high end, that implies a valuation of about $15.6 billion.
So far this quarter, seven companies have postponed or withdrawn their IPOs, compared with just four in the second quarter.
Holtec Nuclear withdrew its filing, while Amaero and SB Energy postponed theirs.
Nvidia-backed NScale may even have to change its roadshow schedule.
U.S. IPOs have raised about $146.9 billion this year.
If Anthropic goes ahead, the full-year total could still surpass the 2021 record.
So, to get back to the question you asked.
Will it trigger a global valuation boom in AI and compute?
My view is that this confirms demand on the compute side, but it won’t ignite it.
About 80% of the $518 billion is non-cancellable, effectively locking in prices for demand over the next ten years.
For cloud providers, chipmakers, and data centers, this means order visibility.
But compute stocks have long moved beyond the storytelling stage.
Nvidia’s net income over the past year was $192.88 billion.
Its trailing 12-month P/E ratio is 29.58.
What it lacks is orders, not a story.
This prospectus confirms the amount of compute, not the multiple.
The model side, though, is where the real valuation anchor lies.
Once the 30.8x revenue figure is priced in publicly, it will become the benchmark for private AI valuations around the world.
If it holds up after listing, private-market valuations will get another boost.
If it opens high and then falls below its offering price like SpaceX, the signal it sends will be a discount.
The test is actually simple.
Watch to see whether it can hold above its offering price in its first month of trading.
China is pointing in the opposite direction.
If DeepSeek files using full-year financial statements, it will bring the valuation multiple for AI stocks in China’s A-share market back to auditable revenue.
The Hong Kong and mainland China markets already have a queue forming.
Moonshot AI has confidentially filed for a Hong Kong listing at a pre-money valuation of $50 billion, while Baidu’s Kunlunxin is pursuing a dual listing.
Plus, after Unitree went public in August, its market cap was about 210 billion yuan.
China will develop its own AI listing sector.
The two markets will diverge rather than rise together.
Because their pricing methodologies and pools of capital are fundamentally separate.
Next, watch for three things that can be verified.
At the October 14 investor day, watch to see whether it gives a price range.
Watch where the November roadshow pricing lands relative to the $2 trillion mark.
How the final prospectus describes the terms and non-cancellable share of the $518 billion commitment will be the blueprint for global compute orders.
Finally, here’s the discipline I’d recommend.
Don’t get in early through secondary-market allocations. Opportunities like those involving eight channels come with governance costs of their own.
Don’t treat the biggest IPO in history as a positive for the AI sector.
Because these two mega-deals are drawing from the same pool of existing capital.
Don’t value compute stocks using revenue multiples.
Nvidia’s multiple is based on net income; the two figures aren’t directly comparable.
One last thing.
Pricing power in this AI cycle is shifting from private-market valuation stories to public-market cash-flow statements.
—MK, Keeping Promises
