Pizi Ge’s Crypto + AI Morning Brief · 2026-10-25

【Crypto】

1. Treasury companies bought 136,000 BTC using preferred stock

According to Fidelity data, from the start of 2025 through Oct. 5, Strategy, Strive and others used preferred stock financing to buy 135,979 BTC—more than 0.5% of the total supply. The preferred stock issued totaled about $12 billion. Last week, Strategy spent $193 million buying back STRC, while purchasing only 267 BTC during the same period. It raised its buyback authorization to $3 billion.

→ The money for buying coins is borrowed, and dividends are a hard expense. Buybacks and coin purchases are competing for the same cash. To assess this chain, watch whether preferred stock is still trading near par: if it falls below par, no one will buy new issuance, leaving buybacks as the only option.

2. X100 hit by a governance attack; about $4.85 million cashed out

The attacker took control of 2 of the 3 Senate seats and sold 400,000 X from the treasury into a self-created empty pool; a second attacker destroyed Freezer’s Uniswap V3 LP. The two transactions cashed out about 4.85 million USDT, and the attack is still ongoing.

→ This wasn’t a code vulnerability, but overly concentrated voting power: controlling 2 of 3 seats is enough to move the treasury, and creating your own pool means setting the execution price yourself. When evaluating a protocol, first count how many votes it takes to pass a proposal.

3. Coinbase premium negative for 31 consecutive days

Latest: -0.1353%; US spot prices remain below Binance. It briefly turned positive at +0.0052% on 8/24, ending a record 97-day negative premium.

→ US buying demand is weak, but don’t read this as institutions pulling out—it tracks only the price spread, not flows. The signal is the day it turns positive, not how long it stays negative.

【AI】

4. DeepSeek open-sources DeepInfer: $38 per GPU-hour on a single card

The cluster processes 155 trillion tokens per day. At off-peak rates, that works out to about $38 per GPU-hour, or $330,000 running at full capacity for a year. The company notes that this is theoretical billable revenue.

→ The word “theoretical” is key: utilization, free credits, and depreciation still need to be factored in. CoreWeave’s second-quarter revenue of $2.575 billion and net loss of $626 million are a useful comparison—cheap tokens don’t necessarily mean profits.

5. AWS computing capacity sold through 2028

Jassy: capacity for 2026 is sold out, 2027 is essentially sold out, and most of 2028 is booked. But capital spending is still committed only a few months ahead; they aren’t locking in for the long term.

→ Demand stretches to 2028, while supply deliberately keeps an escape hatch. They trust orders, not the cycle. In-house Trainium is the lever for cutting costs.

In short: where the money comes from, who has the power to move it, and which way buyers are leaning matter more than the price itself.