Everyone is selling you a solution. No one is showing you the failure mode.
Last week a single line crossed the wire and then vanished into the noise of a bull market: Anthropic, according to the Wall Street Journal, is preparing to commit more than 600 billion dollars to compute. That is the entire signal. One number. No time horizon. No breakdown. No confirmation from the company itself. The original report, as far as I can reconstruct it, ran to roughly a hundred words of secondary summary โ enough to trigger a price reaction, not enough to survive an audit.
I have spent twenty-four years in this industry watching numbers like this arrive, and I have learned that the size of a number is usually inversely proportional to the quality of the thinking behind it. So let me do what I do: strip away the pitch, and look at the protocol. Because a 600-billion-dollar figure that cannot be decomposed is not information. It is a signal. And signals are designed to be read, not believed.
The first thing any auditor does with a number this large is refuse to compare it to anything. You cannot put 600 billion dollars next to OpenAI's 1.4 trillion in infrastructure pledges, or Microsoft's annual capital expenditure, until you know what you are actually holding. Is this a one-year cash outlay? A five-year plan? A fifteen-year commitment? The difference between those three readings is not a rounding error โ it is the difference between a company attempting something physically impossible and a company doing something merely enormous.
So let me anchor it in physics, because physics does not negotiate.
The all-in cost of building an AI data center today โ shell, power infrastructure, cooling, and the IT equipment itself, meaning GPUs, TPUs, networking, storage โ sits somewhere between 30 and 50 billion dollars per gigawatt of IT load. The building and power side runs roughly 10 to 15 billion per gigawatt. The silicon runs 20 to 35 billion, and that figure drifts with every chip generation. Take 60 billion dollars and divide by 40 billion per gigawatt, and you land at roughly 15 gigawatts of IT load.
Now hold that number against reality. The entire United States data center fleet consumed about 176 terawatt-hours in 2023, according to Lawrence Berkeley National Laboratory โ an average load of roughly 20 gigawatts. Even under the most aggressive high-growth scenario for 2028, the whole country reaches something like 66 gigawatts. A single model developer claiming 15 gigawatts would be claiming between 18 and 27 percent of the entire nation's projected data center load. No single company occupies that share of national infrastructure. It does not happen. Not in three years. Not in five.
The only reading that survives contact with physics is that 600 billion dollars is a cumulative multi-year commitment โ almost certainly spanning eight to fifteen years โ and that a large fraction of it lives on somebody else's balance sheet.
That distinction is the entire story, and the report did not contain it.
Consider what happens when you amortize. Spread over ten years, 600 billion is 60 billion a year. That is roughly 70 percent of Microsoft's entire fiscal 2025 capital expenditure, including finance leases. Substantial, but survivable for a company with Anthropic's backing. Spread over five years, it becomes 120 billion annually โ more than any single hyperscaler spends in a year on everything. That reading is not financially coherent. Ten years and beyond is the only span that holds.
And then there is the second silence: the structure. Anthropic does not build like OpenAI. OpenAI has been moving toward vertical integration โ Oracle, Broadcom, custom silicon, owned capacity. Anthropic's compute posture is a three-cloud hybrid: Google TPUs, AWS Trainium, and Microsoft Azure with NVIDIA. Under that architecture, a very large share of the 600 billion is not Anthropic's fixed asset at all. It is a purchase commitment โ a contractual obligation to consume a minimum volume of compute over a defined period, often with a true-up clause if the minimum is not met.
This matters more than the headline. A commitment is not an expenditure. It is a rigidity. It locks in future cost whether or not future demand arrives. In accounting terms it frequently sits off-balance-sheet as a disclosed obligation. In operational terms, it converts demand risk into a contractual liability. If Anthropic's inference demand grows the way the commitment assumes, the structure is brilliant โ volume discounts on TPU and GPU allocations for large buyers can run 40 to 60 percent, meaning a nominal 600 billion may purchase 300 to 400 billion worth of list-price compute. If demand does not arrive, the same structure becomes a sinking cost with penalty exposure.
I have seen this shape before. In 2000, telecom carriers signed long-term capacity commitments on the assumption that internet traffic would double every hundred days. The fiber was real. The commitments were real. The demand curve was not. The assets were not wasted โ they were simply repriced, and the companies that had signed the rigid contracts did not survive the repricing. I am not saying Anthropic is Pets.com. I am saying the instrument has a history, and the report described none of it.
Now let me do something the original piece could not do, because it lacked the numbers: read the commitment as a financing instrument.
Anthropic's annualized revenue, by the best public estimates, sits in the range of 7 to 9 billion dollars as of late 2025. Against that, 600 billion is somewhere between 70 and 85 years of current revenue. Even if revenue triples to 25 billion in 2026, you are still looking at roughly 24 years of revenue. No traditional valuation framework survives that arithmetic. So the framework must be something else.
It is. And the something else is worth naming precisely.
In September 2025, Anthropic closed a 13 billion dollar Series F at a reported 183 billion valuation. Shortly after, market chatter placed a new round near 350 billion. Around the same period, Microsoft and NVIDIA announced investments in Anthropic of up to 5 billion and 10 billion dollars respectively โ and Anthropic committed to purchasing 30 billion dollars of Azure compute. Read that sequence slowly, because the order matters. Investment flows in. Purchase commitment flows out. Revenue gets recognized somewhere in between.
This is the circular structure that has drawn the most scrutiny across the entire AI infrastructure narrative, and Anthropic's 600 billion sits at its center. Google invests in Anthropic, Anthropic buys TPUs. Microsoft and NVIDIA invest in Anthropic, Anthropic buys Azure and NVIDIA silicon. The capital cycles through investment, procurement, and revenue recognition, and each loop makes the next valuation easier to justify. I am not alleging impropriety โ these are disclosed, contractual, auditable arrangements. I am describing a capital structure whose fragility is invisible in a bull market and brutally visible in a correction.
The commitment is not the consequence of the fundraising. It is the precondition for it.
That is the inversion the report missed completely. Publishing a 600-billion-dollar figure is not a disclosure. It is a sales document. It tells a prospective investor: we can lock compute at a scale no competitor can match, therefore the option on our future is worth more than our present revenue. The timing of the leak relative to the fundraising cycle is not a coincidence. It is a mechanism.
And here is where I have to say something that will not be popular in this market. Silence is the loudest audit. The report said nothing about electricity, siting, construction timelines, supply chain constraints, or grid interconnection. That silence tells you what kind of document it is. A capital narrative. Not an engineering feasibility analysis.
Because the real bottleneck is not silicon. It is power.
Fifteen gigawatts of IT load requires fifteen to twenty years of incremental generation and transmission capacity to be built and energized. The interconnection queue in the United States already runs five to seven years in most major markets. Transformer lead times stretch two to three years. High-voltage switchgear is scarce. If Anthropic intends to energize anything close to the implied capacity, the binding constraint is a power purchase agreement, not a purchase order for accelerators. Anyone who has actually walked a data center site knows this. The pitch decks never mention it.
Which brings me to the commercial assumption underneath the number, and this is the part I find most intellectually honest to examine, because it is where the plan either works or does not.
Anthropic's revenue has grown at a rate that is genuinely rare in software history โ from roughly 1 billion in early 2025 to somewhere around 7 to 9 billion annualized by year end. Call it seven to nine times in twelve months. If that slope held, 2027 brings 50 to 70 billion, and 2028 brings 300 to 500 billion. Sustaining a five-times growth rate for three consecutive years has no precedent in enterprise software. Not one.
And the growth is happening against a pricing backdrop that moves the wrong direction. Unit inference cost fell roughly tenfold between 2024 and 2025. If that trend continues, Anthropic would need to increase actual token consumption by more than a hundredfold to hit the revenue implied by the commitment โ while per-token prices fall 50 to 70 percent annually. That is not a marketing problem. That is a physics problem of a different kind, and it requires agentic workloads to move from assisting humans to running autonomously for extended periods.
Which is what the commitment is actually a bet on. Not chatbots. Not even code completion. The 600 billion is an option on the agent economy โ on the assumption that every knowledge worker will eventually run ten to a hundred persistent agents, each consuming tokens continuously. If that world arrives, the commitment looks prescient. If agents remain supervised tools that require a human in the loop, the demand curve never reaches the commitment, and the structure becomes the most expensive miscalculation in the industry's short history.
Let me test this against the competitive picture, because a commitment of this size removes the middle path.
Anthropic's differentiation has never been scale for its own sake. It has been enterprise reliability, code capability, and safety alignment โ a positioning that shows up clearly in Claude Code, which is arguably the highest-value monetization surface in AI today: high seat value, dense token consumption, quantifiable ROI. That is a real moat. It is also a narrow one, concentrated in technical industries where cross-industry penetration remains unproven.
Against that, look at the field. OpenAI's announced infrastructure commitments run to roughly 1.4 trillion dollars over a decade or more. Google self-supplies through TPUs with no external commitment at all. Meta builds its own at roughly 70 billion a year and rising. Microsoft, Amazon, and Google each run annual capital expenditure in the 90 to 125 billion range. Anthropic at 600 billion over ten years is 60 billion a year โ which places a model company's capital intensity within striking distance of a hyperscaler's. That is a category change. Anthropic is no longer positioning as a model developer. It is positioning as a compute holder.
And here is the structural vulnerability that no one wants to name: Anthropic's largest compute supplier is also one of its largest shareholders, and also its most direct competitor. Google supplies TPUs. Google invested. Google's Gemini competes for the same enterprise contracts. The three-cloud neutrality that Anthropic markets is a neutrality of infrastructure, not of capital. A 600-billion-dollar commitment deepens precisely the dependency that makes the position fragile.
There is also a cost dimension the bull case consistently underweights. A multi-cloud hybrid gives Anthropic resilience against any single supplier's failure, but it forfeits the unit-cost advantage of vertical integration. My estimate, based on how custom and merchant silicon economics typically diverge at scale, is that Anthropic's effective cost per unit of compute runs 20 to 40 percent above a fully integrated self-build. In a market where open-weight models are closing on frontier capability โ DeepSeek demonstrated that architectural and training efficiency can cut costs by an order of magnitude โ a heavy-asset depreciation structure becomes a liability rather than a moat. If an open model reaches 90 percent of today's frontier capability by 2027, enterprise switching costs collapse, and the reliability premium compresses fast.
The bull case, to be fair, has an answer. Scale buys price leverage. A 600-billion-dollar commitment is a negotiating instrument, not just a spending plan. It is also a deterrent signal โ to customers deciding which vendor will still be standing in 2029, to researchers choosing where to build, to investors deciding which cap table to join. In a market where perception compounds, the narrative value of the number may exceed its financial rationality. That is a real effect. I do not dismiss it. I simply insist on pricing it as what it is: a marketing instrument with a balance-sheet consequence.
What I refuse to do is pretend the consequence is small.
Run the depreciation. If 600 billion in compute assets depreciates over five to six years, annual depreciation lands between 100 and 120 billion. Add power, operations, and headcount at 20 to 30 percent, and you are carrying 130 to 150 billion in annual cost. To hit a 60 percent gross margin target, you need 320 to 375 billion in annual revenue to cover it. That is more than ten times the revenue the market expects in 2026. The gap between the required number and the projected number is not a forecast error. It is a chasm, and it is bridged by exactly one thing: continued access to external capital at escalating valuations.
Which means the 600 billion commitment quietly sets a deadline. Anthropic must move toward a public offering or a large-scale acquisition before 2027 or 2028, or the financing obligations embedded in the commitment become impossible to service on the assumed schedule. The plan does not describe a business trajectory. It describes a timetable.
I want to be careful here, because caution is the whole point of this exercise. Every number I have used for revenue, valuation, and investment is drawn from public reporting and carries the usual uncertainty. The unit cost figures are industry ranges, not audited figures. The cost-derived projections rest on assumptions about depreciation schedules and margin targets that Anthropic has never disclosed. My confidence in the core structural judgment โ that this is a multi-year commitment rather than annual cash expenditure, and that internal cash flow cannot support it โ is reasonably high, because the physics and the arithmetic both point the same direction. My confidence in any specific quantitative projection is low, and I will say so plainly rather than dress inference up as fact.
That is what an audit sounds like. It is less exciting than a headline.
And now the part that keeps me up at night, which has nothing to do with Anthropic's balance sheet and everything to do with what this class of commitment does to the industry.
When OpenAI commits 1.4 trillion, Anthropic commits 600 billion, Meta commits 70 billion a year, and every hyperscaler expands simultaneously, the aggregate assumption is that demand grows monotonically. History disagrees. Telecommunications fiber in 2000. Shale gas in 2015. Both cycles featured real assets, real contracts, real demand growth โ and still ended in oversupply and repricing, because the commitments were signed against a curve that assumed continuation. If AI compute supply overshoots demand in 2027 or 2028, the correction will not be a pricing event. It will be an asset impairment event, and the companies holding the most rigid commitments will absorb the most damage.
There is a second-order effect that receives almost no attention. A 600-billion-dollar compute program accelerates the automation of exactly the work that funds it. Synthetic data and reinforcement learning environments replace human annotation. That is the direction the capital points. The industry is building the infrastructure that makes a category of knowledge work economically obsolete, and it is financing that construction with the revenue from that work. I do not say this as a moral judgment. I say it as a structural observation, because a system that consumes its own foundation is a system that needs to be understood before it needs to be celebrated.
So where does this leave the reader who is currently, as the market intends, feeling a strong urge to participate?
Three things worth holding onto.
First, stop comparing the number to other numbers. A commitment without a disclosed time horizon and composition is not a comparable figure. Anyone who places 600 billion next to 1.4 trillion and draws a conclusion is committing a dimensional error, and dimensional errors are how portfolios die.
Second, watch the electricity, not the silicon. If Anthropic has signed nuclear or natural gas power purchase agreements at gigawatt scale, the commitment is real and the timeline is credible. If the announcements keep coming about chips and stay silent about interconnection queues, transformers, and water rights, then you are watching a financing narrative, and you should price it accordingly. The power contract is the only document that cannot be faked.
Third, remember what a commitment actually is. It is a promise about a future you do not control, secured against a present you do control. The people signing these documents understand that. The people reading the headline usually do not.
I have been in this industry long enough to have watched the 2017 mania, the 2020 yield farms, the 2022 collapse, and the institutional arrival that followed. The pattern never changes. The technology improves. The commitments grow. The failure modes stay exactly where they have always been โ in the gap between what is promised and what is verified. What changes is only the size of the number we are asked to accept without evidence.
Code is law only when it is aligned with human values. A number is truth only when it survives decomposition. Anthropic has given us a number, and the market has given us a reaction, and neither has given us an audit. Until someone produces the breakdown โ the time horizon, the structure, the power contracts, the funding sources โ what we have is not a fact about the future of compute.
It is a pitch. And the protocol is still the only thing worth trusting.
The question I leave you with is not whether Anthropic can spend 600 billion dollars. It is whether an industry that has learned to price promises this large has retained any capacity to verify them at all โ or whether we have quietly agreed that scale itself is the only evidence we still require.