Everything About AI
NextFin News - Anthropic, the maker of Claude, is preparing to spend as much as $750 billion on additional computing power in the coming years, a scale of capital commitment that would rank among the largest corporate investment plans ever disclosed, as the artificial-intelligence company readies an initial public offering. The figure sits inside a draft IPO prospectus filed confidentially with the Securities and Exchange Commission, and it arrives alongside a starker number: a net loss of $80.7 billion in the first half of 2026, up from $13 billion in the same period a year ago.
The two figures together frame the central tension of the most scrutinized technology listing in years. Anthropic is betting that frontier AI is won by whoever can lock down the most compute, fastest — and that revenue growth will outrun a loss curve that has steepened more than sixfold in twelve months. Investors who buy the stock will be buying into a business model that inverts the traditional software promise: instead of capital-light margins, they get capital-heavy spending that must be prepaid before the customers arrive.
The Filing in Numbers: A Sixfold Loss Widening
The draft prospectus lays out a company growing at a pace that is difficult to compare with anything in public software markets. Revenue run rate climbed from roughly $14 billion in February 2026 to about $47 billion by mid-May, and first-half 2026 revenue reached approximately $9.6 billion. The company told investors earlier this year that it projected $10.9 billion of revenue in the second quarter alone — more than double the $4.8 billion posted in the first quarter — and expected its first quarterly operating profit of about $559 million.
Against that growth, the losses are the headline shock. An $80.7 billion net loss in six months implies a burn rate that exceeds the annual revenue of nearly every software company in the S&P 500. The year-over-year comparison is the real story: the loss in the first half of 2026 was more than six times the $13 billion recorded in the first half of 2025. That is not a one-time charge pattern; it is the operating profile of a company spending ahead of demand on a scale that has no clean public-market analog.
The valuation math has raced even faster. Anthropic's private-market value moved from about $183 billion in March 2026 to roughly $965 billion in May, with internal targets pointing toward $2 trillion. A term sheet reported in the spring would have placed the company at a $930 billion post-money valuation. At that level, the IPO would become the largest software listing in history — and the market would be asked to underwrite a balance sheet whose defining feature is a multi-hundred-billion-dollar compute obligation.
That obligation is already taking shape. Ahead of the filing, the company had locked down approximately $517 billion in compute commitments, and the new $750 billion figure points to still more capacity on top of what is already contracted — a cumulative compute bill that, if realized, would exceed the annual economic output of most countries and dwarf the capital-expenditure plans of the world's largest cloud providers.
Why This Is Structural, Not Cyclical
The first question any investor should ask is whether this spending is a cyclical build-out — a temporary capacity grab that will normalize once the infrastructure is installed — or a structural change in what it costs to compete in AI. The answer, on the evidence in the filing, is structural, and it matters because the two require completely different valuation frameworks.
A cyclical build-out has three markers: a short-term demand spike, a finite installation period, and a return to a steady-state capital intensity once the build is complete. Think of a semiconductor fab cycle or a telecom network rollout: heavy spending for two or three years, then depreciation and maintenance. Frontier AI does not fit that pattern. The models do not stop improving, the competitive bar does not stop rising, and the compute required to train the next generation grows faster than the efficiency gains from better chips. Every dollar of capacity Anthropic installs today is obsolete for training purposes by the time the next model generation ships. The spending is not a project; it is the product.
There is also a history that no longer applies. For two decades, software investors were trained to reward capital efficiency: high gross margins, low capex, operating leverage that flows straight to the bottom line. That rule set assumed code could be replicated at near-zero marginal cost. Frontier AI breaks the assumption. Inference — running the model for each user query — carries a real and recurring compute cost per token, and training the next model requires clusters so large that only a handful of entities on Earth can assemble them. The marginal cost of software has moved from effectively zero to something that scales with usage. That is a regime shift, not a cycle, and it will not mean-revert on its own.
The strongest evidence that this is structural sits in the competitive dynamics. This is not one company making a one-time bet; it is an arms race with at least four well-capitalized participants. OpenAI, Anthropic's direct rival, is itself loss-making and racing toward its own public listing. The hyperscale cloud providers — Amazon, Google, Microsoft — are spending tens of billions annually on AI infrastructure, with one of them guiding capital expenditure toward the $130 billion to $145 billion range for the year. The chip supplier at the center of the stack reported quarterly revenue of $96 billion, up 106 percent year over year, and guided the following quarter to $108 billion. When every major player is committing hundreds of billions simultaneously, no single company can unilaterally de-escalate without conceding model leadership. The spending is a prisoner's dilemma, and the equilibrium is more spending, not less.
The Second-Order Trade: Who Actually Captures the Value
The first-order read of the filing is straightforward: Anthropic is spending enormous sums, losing enormous sums, and the IPO prices that tension. The second-order question is more important, and the market has not fully priced it: in a compute arms race, the most reliable profits sit upstream, not at the model layer.
Follow the dollar. For every dollar Anthropic commits to compute, a large share flows to the chip designer, the foundry, the power and cooling infrastructure, and the cloud provider that finances and hosts the capacity. Anthropic's own gross margin must first cover that bill before it reaches operating profit. The suppliers, by contrast, get paid as the capacity is ordered, regardless of whether Anthropic's models win the next benchmark. This is the economic logic that has made the AI infrastructure layer the cleanest trade of the cycle: demand is inelastic, pricing power is concentrated, and the customer's urgency is the supplier's margin.
The exposed party is the public-market investor in the model layer itself. Private markets can tolerate loss math that public markets scrutinize quarterly. In the private rounds, investors underwrote Anthropic at $965 billion on the strength of a revenue trajectory that had not yet been audited, a customer base concentrated in API and enterprise contracts, and a safety narrative that doubled as a moat. The public market will see the same numbers under daylight, with a different incentive structure: quarterly earnings, shareholder votes, and a share price that reacts to every miss. The gap between what private capital priced and what public capital will accept is the IPO's central risk.
There is also a timing asymmetry that favors the suppliers. Anthropic's compute commitments are largely prepaid or take-or-pay obligations — capacity that must be paid for whether or not it is fully utilized. Utilization is the swing variable. If demand keeps compounding at the current pace, the capacity gets absorbed and the losses compress toward profitability. If demand growth decelerates even modestly, the company is left paying for idle GPUs while competitors keep building. The fixed-cost base, once installed, cannot be turned off.
Investors are paying roughly 1.8 times more per equity dollar deployed into Anthropic, and that gap has been widening for two consecutive rounds.
Harrison Rolfes, a senior research analyst at PitchBook, made that assessment, and it captures the premium the market has already assigned to Anthropic's growth. It also frames the burden of proof: the company must not merely grow, but grow fast enough to justify a per-dollar valuation well above its closest rival.
The Counter-Thesis: Growth Can Outrun the Losses
The bear case is seductive, but it has a serious flaw, and the bull case deserves its full weight. The strongest argument for Anthropic is that the losses are a timing artifact of a revenue curve that is still in its steepest phase. If revenue doubles every few months, today's $80.7 billion loss is a small fraction of a much larger future income stream. The company projected its first operating profit — $559 million — within the year. Once fixed capacity is installed, each incremental dollar of revenue requires less incremental spend, and operating leverage kicks in hard.
There is evidence that demand is real, not speculative. A corporate spending tracker reported in May 2026 that 34.4 percent of U.S. businesses had adopted Anthropic's tools, compared with 32.3 percent for OpenAI — the first crossover on record. That is enterprise adoption, not retail novelty, and it suggests the revenue base has a durability that a chatbot fad would not support. If Anthropic holds that adoption lead and monetizes it through API usage and enterprise contracts, the path from $9.6 billion in half-year revenue to a $100 billion annualized run rate is not fantasy; it is a continuation of a trend already in motion.
The counter-thesis also has a point about pricing power. As models become more capable, customers pay for outcomes, not tokens — and outcomes command premium pricing. A model that replaces a team of engineers or automates a workflow can be priced as a fraction of the labor it displaces, which leaves wide margin room even after compute costs. In that world, the $750 billion spend is not a burden; it is the price of admission to the highest-margin software market ever created.
But the bull case rests on one assumption that must hold: utilization. The entire model depends on the installed capacity being used. If utilization of committed capacity falls materially — say, below 60 percent — the fixed-cost base becomes a drag that no amount of pricing power can fully offset. That is the specific signal that would break the structural-bet thesis.
What to Watch: Three Horizons, Three Scenarios
Short term (the IPO window): The listing is expected to be the most scrutinized technology offering in years, with the prospectus still moving through the confidential review process after a June 1 submission. The immediate risk is pricing: if the company holds firm on a valuation near the $2 trillion target while disclosing an $80.7 billion half-year loss, the order book could thin quickly. A more modest debut would leave room for a post-listing rally; an aggressive one invites a first-day reset.
Medium term (fundamentals, 12–24 months): Watch three metrics. First, the revenue run rate: the company must demonstrate a credible path to roughly $100 billion annualized within a year of listing, or the growth premium evaporates. Second, gross margin on inference: after recent price cuts on its top-tier model, does margin expand with utilization or compress under competitive pressure? Third, utilization of committed capacity — the single number that separates a growth story from a stranded-asset story.
Long term (structural, 3+ years): The question is whether frontier AI settles into an oligopoly with durable margins, or whether it becomes a commoditized utility where the value accrues to the infrastructure owners. If Anthropic can maintain a model-quality lead and convert enterprise adoption into multi-year contracts, it can defend pricing. If model capabilities converge and customers shop on price, the $750 billion compute bill becomes a tollbooth to a low-margin business.
The base case is a volatile listing followed by a two-year proof period in which revenue growth either validates the private-market valuation or forces a painful repricing. The upside case is that adoption keeps compounding, utilization stays above 70 percent, and Anthropic reaches sustained operating profitability, becoming the most valuable software company in history. The downside case is a growth deceleration that exposes the fixed-cost base: revenue misses the $100 billion threshold, utilization falls toward 50 percent, and the stock trades down to a multiple of tangible capacity rather than growth optionality.
The specific falsifying signal is quantifiable: if Anthropic's disclosed revenue run rate fails to approach $100 billion annualized within 12 months of the IPO, or if committed-capacity utilization drops below 60 percent, the thesis that this spending is a structural winner-take-most bet is wrong — and the market will reprice the stock as a capital-intensive utility, not a software franchise.
The Bottom Line
Anthropic's $750 billion compute commitment is not a budget line; it is a declaration of how the AI era will be won. The company has chosen to compete on scale of capital rather than capital efficiency, betting that in a frontier race, the second-place model is worth zero. Investors will soon get to vote on whether that bet is a path to a $2 trillion franchise or the most expensive lesson in software economics ever taught. The filing makes one thing certain: the era of capital-light software is over, and the companies that survive will be the ones that can raise, spend, and deploy hundreds of billions without blinking.
Explore more exclusive insights at nextfin.ai.
