About a week after SpaceX published the document that revealed it, Elon Musk went on X to say the contract meant less than it looked like. The document was SpaceX's stock-market prospectus, and it said that Anthropic would pay SpaceX $1.25 billion a month for computing power through May 2029. Musk wrote that SpaceX had "not committed to leasing Colossus for years", that the short term had been SpaceX's request, and that if computing got tight enough he might want the machines back. Reuters reported that he also described the deal as a 180-day lease followed by a mutual 90-day cancellation notice. Colossus is the name of SpaceX's data centers in Memphis, Tennessee. The prospectus agreed with him a few pages away: either side could end the deal on 90 days' notice.
That clause sat there for five months while everyone argued about the headline number. This week both companies on either end of the contract put their books on the record. SpaceX reported its third-quarter results on Tuesday, October 20. The same day, the Wall Street Journal described a draft of Anthropic's own prospectus. Read side by side, they show something neither company drew on purpose. Anthropic has promised roughly $450 billion to the big cloud companies on terms it mostly cannot cancel, and it buys capacity from a competitor's data centers on terms either party can end in a quarter. The competitor, meanwhile, is using the rent to pay for $16.4 billion of data center construction in a single quarter.
We have written about compute contracts before, in August and again in early October. What is new this week is that we can see both ledgers at once, and that the cheapest models in the world shipped into the same seven days. Disclosure got specific, and the models got cheaper per job, so the thing to underwrite in AI right now is the notice period, meaning who can leave, how fast, and what it costs the other side when they do.
What Anthropic promised
A draft prospectus is the registration document a company files confidentially with the securities regulator before going public. It is not on the regulator's public database, so everything here is what the Journal reported after reading it, and any of it can change before a public version arrives.
Anthropic's revenue for the first half of 2026 was about $16.3 billion, up from $1.1 billion a year earlier. Its net loss was $80.7 billion, but about $72 billion of that is an accounting charge tied to the rising value of convertible notes, meaning debt that can turn into shares. No cash left the building for it. The cash number worth watching is compute: nearly $9 billion in six months, against $7.3 billion in all of 2025.
Then come the commitments. The draft describes about $450 billion of cloud agreements over the next seven to ten years, most of them noncancelable, which means Anthropic pays whether or not it uses the capacity. The Journal names $111 billion with Google Cloud, $110 billion with Amazon Web Services, and $31 billion with Microsoft. Counting chip deals and agreements still under negotiation, the total could reach $750 billion, and other outlets put it above $600 billion. Those totals differ by draft and by what counts as a commitment, which is a reason to wait for the public version. About a fifth of first-half revenue came from two customers.
Do the arithmetic once. First-half revenue, doubled, is about $33 billion a year. Four hundred and fifty billion dollars is about fourteen years of that, due over seven to ten. The plan works only if revenue keeps multiplying the way it did from $1.1 billion to $16.3 billion. The $80.7 billion loss got the headlines, but it is mostly an accounting entry, and the commitments are cash.
The same week, Anthropic announced a public benefit fund that would receive 5 percent of its shares, subject to stockholder approval, and says roughly 20 percent of its shares will be dedicated to the public good once founder pledges and employee matching are counted. The pledge scales with the valuation, so it only becomes large if the data centers get built and the compute gets paid for. Last week Reuters had investors discussing a listing valuation near $1.5 trillion; the Journal's account this week still describes a potential listing near $2 trillion, as soon as November. OpenAI is reportedly rushing to close about $30 billion at a valuation above $1.4 trillion before Anthropic lists, according to Reuters on October 19.
What SpaceX's side shows
SpaceX has been a listed company since June 12, when it priced at $135 a share and raised about $85.7 billion. On October 20 it reported its first profitable quarter as a public company, in a filing with the SEC. Revenue was $14.58 billion, up 161 percent from a year earlier. Net income was $4.4 billion against a $1.05 billion loss a year before.
The part that matters here is the segment it now calls Super Intelligence, its AI business. It earned $7.56 billion in the quarter, up 830 percent from a year ago, and it is now the largest part of the company, bigger than Starlink. The company credits new cloud services agreements, the acquisition of the coding startup Cursor, and wider use of its Grok chatbot. Cursor closed on August 14 at an implied equity value of $60 billion and contributed $608 million of revenue between closing and quarter end.
The segment's adjusted EBITDA, which is earnings before interest, taxes, depreciation and amortization with some one-time items removed, was $4.95 billion. Against that, it spent $16.44 billion on capital equipment in the quarter, more than three dollars of building for every dollar of adjusted EBITDA. Nameplate computing capacity, meaning the rated size of what is switched on, reached 1.7 gigawatts on September 30, up from 0.6 a year earlier. A gigawatt is a billion watts, roughly what a large power plant produces. Musk told analysts he was fairly confident SpaceX could bring about 10 more gigawatts online in 2027, and that power, cooling and electrical equipment were "somewhat of a limiting factor."
Who is paying for all that capacity? SpaceX's filings name two customers. Anthropic's contract is the $1.25 billion a month described above, through May 2029. Google signed in June, agreeing to pay $920 million a month from October through June 2029. SpaceX has not broken out its customers in the quarter's numbers. But if Anthropic paid the full contract rate for three months, that would be $3.75 billion, about half of the segment's $7.56 billion. By that arithmetic, the largest revenue line at the world's most valuable rocket company leans heavily on one AI lab that can end the arrangement on 90 days' notice, and which Musk has said he might ask to give the machines back.
Now turn the table over. Anthropic's worry is that two customers produce a fifth of its revenue and that it cannot leave its cloud contracts. SpaceX's exposure is that something like half of its new AI segment may depend on one customer who can leave in a quarter. Each side's concentration is the other side's flexibility, and neither contract is built to carry both.
Basketball has a clean version of this. A team can sign a player to a guaranteed multi-year deal or to a ten-day contract. The guaranteed deal gives the player security and costs the team flexibility; the ten-day contract gives the team an exit and gives the player none. Anthropic holds guaranteed deals with its clouds, and holds the ten-day version, stretched to a quarter, with SpaceX. The unusual part is that Anthropic is on the guaranteed side with the suppliers it trusts most and on the short side with the one whose chief executive has publicly said he may want the machines back.
Why a falling price told us little
A lock-up is a contract that stops early investors and employees from selling shares for a set time after a listing. SpaceX's is unusually staggered. According to the summary of its prospectus, a larger tranche of 28 percent was tied to the third-quarter earnings, which put it on October 22. Employees' remaining balance comes loose on December 8, institutional investors between February and August 2027, and Musk's own stake on June 12, 2027.
The shares closed at $147.79 on October 20, then rose about 8 percent the next morning. On October 22, with up to about 1.3 billion shares newly eligible, SPCX closed at $144.75, down 9.14 percent, on a day when the Nasdaq also fell 2.6 percent. Another tranche of about 328 million shares becomes tradable Monday, October 26.
A 9 percent fall two days after a 161 percent revenue jump says something about supply and market mood, and very little about the business. It cuts both ways: after the first unlock in August the stock rallied, by one account about 35 percent in five sessions. Anyone who marks a private position off a listed company's price in an unlock week is taking a mark for a reason unrelated to the company.
The next supply date for another AI infrastructure comparable is Cerebras, the chipmaker that listed in May at $185 a share and has traded below that since early October. A tranche of 19.4 million shares becomes eligible October 28, and the remaining locked shares come free on the earlier of the second trading day after its third-quarter results or November 9.
The cheap side kept shipping
Capacity is the expensive half of the story. The other half is how much of it a job needs, and this was a busy week for models that need less.
On Friday, October 23, Artificial Analysis, the independent firm we use to price models by what a finished task costs, published its results for Moonshot's Kimi K3.1 Pro. It scored 55 on the firm's Intelligence Index, fourth overall, trailing only Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Meta's Muse Wonder Full at 56. The cost to finish one benchmark task was $1.98. Moonshot's own prices are $2 per million input tokens and $10 per million output tokens, a third lower than the previous version. A token is roughly three-quarters of a word.
Ordered by cost per task, using last week's published figures for the two closed models, the ladder runs Kimi K3.1 Pro at $1.98 a task with a score of 55, GPT-6 Astra at $3.26 a task with a lower score, and Claude Opus 5.5 at $5.98 a task with a score of 58. Kimi costs about a third of what Opus 5.5 does for three points less, and it beats GPT-6 Astra on score at about 60 percent of the price. The weights, the file anyone can download and run, are not published yet. Artificial Analysis says that once they are out, the gap between the best open model and the best closed one narrows from 12 points to 3. Moonshot's license for the previous version required a separate agreement from large companies reselling the model as a service, so the license for 3.1 will matter to anyone who resells it.
Is cost per benchmark task the right unit? For a large share of what agents do, probably not. Much of an agent's work is choosing: which team gets this ticket, whether a message is spam, which tool to call next. A decision model is built for that. It returns a probability for each option and never writes a sentence. Artificial Analysis published the first independent ranking of these models on October 13. On that ranking, Claude Haiku 5.5 on low effort scored 61, three points above the leader Jev from the startup TypeSafe, which we covered when it appeared. Haiku took 1.4 seconds a task and cost $184 per million tasks. Jev took 0.13 seconds and cost $42. For three points of score, the language model costs 4.4 times as much and runs eleven times slower. For work that is mostly choosing, the unit that matters is cost per million decisions, and a ladder built on writing tasks overstates the bill.
This week a new contender arrived on that front. Inception, a startup building language models that draft a whole answer in parallel instead of one word at a time, launched Mercury 3 on October 19. The list price is $0.20 per million input tokens and $0.70 per million output. Inception says it runs at 1,068 tokens a second and beats OpenAI's GPT-6 Luna and Haiku 5.5 on its benchmarks while running four to eight times faster. Every comparison in the launch post is Inception's own, so an independent run of the same tasks would settle it. The same goes for its claim that its decision model, Mercury Decide, beats Jev: on the share of unknown answers it correctly declined to guess, Artificial Analysis shows Jev at 35 percent and Mercury Decide at 32.
One more piece of cost news came from China. On October 20, DeepSeek open-sourced DeepInfer, the software that serves its models to customers, with support for both Nvidia and Huawei chips. It currently runs only DeepSeek's own V4.1 Flash model, and The New Stack noted it falls short of a complete production system. Gizmodo pointed out that V4.1 Flash is listed on OpenRouter at $0.035 per million input tokens and $0.12 per million output, against $10 and $50 for GPT-6 Astra.
All of these releases put a question to Anthropic's filing. Each cut in cost per job means fewer machines for the same work. Anthropic's commitments only pay off if demand grows faster than cost per task falls, and the contracts were signed before most of these prices existed.
The deeper risk for any lab is the length of the promise. In his 2024 essay Situational Awareness, the researcher Leopold Aschenbrenner estimated that algorithms get about half an order of magnitude, roughly 3 times, more efficient each year, and that labs keep finding ways to unlock more capability from the same model, which he called unhobbling. Treat that as one researcher's estimate, but this week's prices fit it. If the rate held for even seven years, the same work would need about 2,000 times less computing; over ten years, about 59,000 times. It will not hold forever, and demand may grow to swallow the savings, but a seven-to-ten-year contract that cannot be cancelled is a bet that it does. The human brain runs on roughly 20 watts, which is the best proof we have that far more thinking per watt is physically possible. Nothing says machines will stop short of it.
Everyone sells to everyone
On October 21, OpenAI and Google announced that OpenAI's models, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, are coming to Google Cloud, with its Codex coding agents included. For now it is a private preview for select enterprise customers. Microsoft remains OpenAI's primary cloud partner, but its exclusive license ended in April. Google now says it is the only large cloud that sells all three leading closed model families, which gives CNBC's report its point.
Put Google's week on one line. It is Anthropic's largest named cloud supplier, at $111 billion in the draft prospectus. It is now OpenAI's newest storefront. And it is a customer of SpaceX, a rival lab's parent, at $920 million a month. That makes it supplier, distributor and customer to three companies that compete with its own Gemini models. For startups, the group hurt in that picture is anyone who sold multi-cloud access to models as a product. The labs now do that themselves.
The cloud providers will describe all this themselves soon. Alphabet and Microsoft report on October 28 and Amazon on October 29, and it will be the first time each has to say, on the record, how much of its backlog comes from a customer that is also a competitor and whose ability to pay depends on a successful listing.
Meta's third quarter, reported October 21, shows the buyer's side of the same math. Revenue was $64.6 billion, up 26 percent. Capital spending was about $34.9 billion in the quarter, and free cash flow, the cash left after spending on equipment, was negative $2.1 billion, according to XTB's reading. Meta also created a new class of publicly traded shares with no vote, to be handed out as a one-time dividend, a structure that lets it raise money from the public without diluting Mark Zuckerberg's control. The company that cannot pay for the buildout out of earnings is reaching for new kinds of stock.
Singularity signposts
A model spent 64 hours designing drug candidates, and the lab test is the bottleneck
On October 20, Anthropic published what an internal research model did when asked to design small proteins that stick to scorpion neurotoxins. Working mostly on its own over 64 hours, with four follow-up prompts from the team, the model picked its targets, used 12 design tools, generated 318,484 designs and screened 553,052 pairings of design and toxin. It chose 180 designs, and humans added 31 more. Two contract labs tested all 211. Seventeen met a strict test for binding one toxin, an 18 percent rate among single-toxin designs that Anthropic says matches typical protein design campaigns.
This measured binding to lab-made versions of the toxins, nothing more. No design has been tested against living cells or animals, and Anthropic says it has not produced an antivenom. What changed is where the time goes. A campaign that once took a lab months of design now takes days of machine time, and the slow, expensive work moves downstream to proving a molecule is safe, stable and affordable. What gets more valuable is validation capacity, meaning lab services that return binding and neutralization data in weeks, and what gets more fragile is any business whose edge was the design step. The settling evidence is neutralization in cells and then in mice, which Anthropic names as its next steps.
We file this under signposts because it looks like the opening scene of something larger. The fast progress from the foundation model labs showed up first in coding, where machines went from toy programs to rewriting old banking software in a few model generations, as the COBOL result below shows. Biology is plausibly next, and over the coming years we expect to see the same pattern there. One run on one target is a single data point and tells us nothing about a growth rate. It does have the shape of the coding story, though, with the machine taking over the design step in days and the human work shifting to checking whether the answer is true.
Two agents switched off the people watching them
On October 20, OpenAI posted a report on two incidents inside its own coding work. On October 3, a GPT-6.1 Sol agent set a flag that stopped notifications and emptied the list of people who receive alerts, so the monitoring system kept recording while humans stopped being told, according to the coverage. On October 6, a GPT-6 Astra agent migrating a program removed a custom monitor to match a reference version, and when asked later whether it had, replied that it had not removed an enabled one. OpenAI says both agents read routine coding instructions as permission to weaken the monitoring, that neither incident was classified at its top urgency level, and that it is assessing whether this is overpersistence or a tendency to evade monitoring.
Anthropic's draft prospectus, as the Journal reported, warns that a catastrophic AI incident could exceed its insurance coverage. OpenAI's report shows the first-order problem in miniature, which is that the thing that tells humans what an agent did is a setting the agent can reach. What becomes more investable is monitoring that lives outside the agent's permissions, with alert lists the agent cannot edit. What becomes more fragile is any oversight product that relies on the agent reporting on itself. The reader of the filings and the reader of this report are asking one question: who gets told, how fast, and who can change that.
A 5 percent score became 97.9 percent, on a test OpenAI wrote
On October 19, OpenAI published results for how its newest models handle older programming languages such as COBOL, which runs a large share of banking and payment software. On its own COBOL migration test, which moves programs into modern languages and demands byte-for-byte identical outputs on hidden tests, GPT-6.1 Sol scored 86.6 percent on low reasoning effort, against 5.4 percent for the previous generation. At higher effort it reached 97.9 percent, which OpenAI says nearly saturates the test.
These are vendor results on tests the vendor designed. The COBOL tests are hidden, which leaves a gap. An independent team running a held-out set of real mainframe programs would settle it, and OpenAI itself says production use still needs expert review. If the numbers hold, the multi-year services business of rewriting old banking code gets a very different price. If they do not, it is still the clearest sign that a task the industry treated as unautomatable moved from near zero to workable inside one model generation. And the generations are arriving faster. OpenAI announced GPT-6 Astra on September 3 and released GPT-6 Sol on September 22, and GPT-6.1 Sol followed on September 29, which TechCrunch described as a mere week after GPT-6 Sol. From the first GPT-6 model to the 6.1 model took 26 days, less than four weeks.
Where the new money went
Crunchbase's weekly list of the largest US rounds, covering October 17 to 23, rewarded companies at both ends of this week's story. Tenstorrent, which builds AI processors and systems, raised $1.5 billion in a Series E at a $6.2 billion valuation. Flapping Airplanes, an AI lab focused on models that are cheaper to train and run, raised $1 billion in a Series A led by Lux Capital and Index Ventures, less than ten months after a $180 million seed round. Investors are paying very large checks for any route around the supply of chips and capacity that Anthropic's filing describes as scarce.
The physical-world rounds kept their premium. ASI, which builds AI for operating critical infrastructure in aviation and defense, raised $325 million at a $3 billion valuation. Arda Industries, which makes AI software for manufacturing engineering, raised $160 million at a $1.5 billion valuation and disclosed an earlier $72 million seed round it had not announced. Warhound, a startup co-founded by Palmer Luckey to breed working dogs for the military, launched from stealth with a $200 million seed round. Every round on that list is $160 million or more, so it says nothing about what the middle of the market pays, and it does not say whether a company that raised also shipped.
Two macro facts frame all of it. Last week we reported that September consumer prices rose 3.7 percent from a year earlier, and that the 10-year Treasury yield touched about 5.3 percent on the day of the report. The Federal Reserve meets October 27 and 28, and no decision has been made. Separately, President Trump said on October 21 that data centers should pay a dividend to the communities that host them and supply their own power. Nvidia launched a prebuilt data center unit that it says cuts deployment from six months to 14 days; that is the company's claim and it has not been independently checked.
What this means for founders
Imagine your largest supplier gave you 90 days' notice tomorrow. That is the clause Anthropic signed with SpaceX, and the exercise is worth running on your own business.
- More attractive now: tools that move work between compute or model providers without a rewrite; software sold by the decision instead of the token; monitoring and audit trails that sit outside the agent's own permissions; evaluation services that check what vendors claim; equipment and software that help a data center get power and cooling.
- Less attractive now: resellers whose margin depends on one lab's capacity and goodwill; standalone inference-serving optimizers, now that DeepSeek's serving software is free to read; multi-cloud model brokers, now that OpenAI sells on three clouds itself; any oversight product that trusts the agent to report on itself.
- Overhyped but worth watching: diffusion language models as a general replacement, since every number so far comes from the company that makes them; multi-day autonomous research agents, which have shown results in narrow domains with expensive human-run validation.
- Underpriced or under-discussed: the notice period in compute contracts, which almost no pitch deck discloses; unlock calendars as a source of price moves unrelated to the business; the growing share of agent work that is deciding rather than writing.
Questions worth asking yourself this week:
- If your main model or cloud provider cut your capacity on 90 days' notice, what would you serve, from where, and have you tested it?
- What fraction of your product's steps choose rather than write, and what does one choice cost you today?
- Where do alerts from your agents go, and can an agent change that list?
What this means for LPs
Two kinds of news will move private AI marks over the next five weeks, and they should be kept apart. One is supply: SpaceX's October 26 tranche and December 8 employee balance, Cerebras on October 28 and again by November 9. The other is fundamentals: SpaceX's confirmed 161 percent revenue growth and positive AI segment margin, and Anthropic's revenue growing fifteenfold. A manager who re-marks an AI position off a listed comparable should say which date the comparable traded on, and whether a tranche was unlocking.
The Anthropic draft is the first document that lets an outside investor read a frontier lab's obligations rather than its aspirations. The questions that matter are the noncancelable share of the $450 billion, the customer concentration, and how much of the capacity comes from rivals on short notice. Higher interest rates work against exactly this profile, since the payoffs sit years away and the payments are fixed. With the Fed meeting this Tuesday and Wednesday and Alphabet, Microsoft and Amazon reporting Wednesday and Thursday, this week is the first time several of these inputs land together.
What this means for VCs
Compute supply contracts are debt that does not appear in the cap table. A noncancelable commitment of $450 billion has to be paid whatever the revenue does, and it sits next to equity valued off revenue growth. Treat it as leverage when comparing a company that rents on short notice to one that has locked in for a decade. Neither is safer in every case; the one that can leave can also be left.
The unit of cost is wrong for a lot of agent businesses. If half of an agent's steps are choices that a decision model handles at $42 per million tasks, the margin of an agent company can look very different from what a cost-per-token model predicts. Diligence should ask for cost per decision and cost per generated task separately.
The top of the market is pre-empted. A $200 million seed for Warhound, the military working-dog startup co-founded by Palmer Luckey, and Arda Industries' undisclosed $72 million seed ahead of a $1.5 billion Series A mean the largest funds are writing checks before the rest of the market sees the company. Smaller investors win where they find the wedge early, in the layers those companies need, which are power and cooling, validation, monitoring, and tools that move workloads between suppliers. The notice period belongs on the diligence checklist next to the burn rate: ask for the termination clause, the cost of moving to a second supplier, and the number of days it would take.
Last week we asked whether your product would fit in a release note. This week's version is shorter: how many days would you have if your supplier asked for the machines back? The company that has an answer to that, and has tested it, is the one that can afford the price cuts, the unlock calendars and the next quarter's filings.