The financial stakes in pacing the frontier

Rising research bills and shrinking task margins reinforce the race.

A lab funding a larger development program to restore its premium invites a competitive response. When rivals answer with their own investments, expenditure rises while successive releases shorten the premium’s life. That mechanism predicts financing pressure when development outgrows total customer contribution. It does not establish that task-price declines translate directly into falling company revenue or that all labs follow the same path.

The essay’s interactive comparison tests different release intervals over 36 months. Model releases are 6, 10.5 and 15 months apart in the three scenarios. Each assumed release cycle funds one portfolio development budget, and each successive budget increases by the same chosen percentage. Task volume, task price and serving cost follow the same assumptions across scenarios. Price protection and lost task volume from pacing are both zero by default and remain editable.

The scenarios isolate the effect of release timing while holding the economic assumptions constant. They have no assigned probabilities and are not forecasts of named labs. Initial scale comes from the same reported revenue, R&D and gross-margin anchors described below. Those anchors do not validate future growth assumptions.

At elapsed time t in years, racing task volume is proportional to (1 + demand growth)^t, price to (1 − annual price decline)^t, and unit serving cost to (1 − annual cost decline)^t. Initial annual serving cost is initial annual revenue × (1 − initial contribution margin). Revenue and serving expense each multiply the applicable unit amount by task volume. Their difference is contribution before development and other costs.

For the paced path, price protection reduces the annual percentage price decline by the chosen fraction; 50% protection turns a 30% annual decline into 15%. The annual task-volume shortfall multiplies volume by (1 − shortfall)^t. A 60% shortfall therefore leaves 40% of the racing volume after one year and 6.4% after three; it is not a one-time 60% haircut. When no delay is selected, both pacing effects are zero. These are explicit counterfactual inputs, not estimated effects of a particular policy.

A selected share of each development budget is paid at its start; the remainder is spread evenly through its cycle. The default is 25% upfront. Contributions use each calendar month’s midpoint rate, split proportionally at a fractional-month boundary. Budgets start before month 36, with none opened exactly at the horizon. The paced settings leave a development round in progress; its remaining payments fall beyond the comparison period and are reported separately. The year-3 coverage measure assigns a deposit at month 24 to year 3; it divides year-3 contribution by that year’s development payments separately for each scenario. It is not the share of all three years’ spending recovered.

Peak uncovered spending is the maximum cumulative development payments less customer contribution, floored at zero. The end balance keeps its sign: negative uncovered spending denotes a surplus before omitted costs. These quantities do not include cash already raised, overhead, tax, interest or working capital. The model permits negative serving contribution rather than hiding it, and omits responses such as changing prices, capacity or strategy. It models chosen paths, not a strategic equilibrium.

The slower lab funds fewer development rounds, so the comparison does not grant equal capabilities for less investment. The cash-flow comparison stops at the horizon, including partial payments on any round still in progress. Unpaid commitments on that round are shown separately; future earnings and new rounds beyond the horizon are not valued. Its end balance is not lifetime profit. Price and demand assumptions partly expose the cost of delaying useful models; public evidence does not identify the full counterfactual.

Download all assumptions and their provenance · Download scenario cash-flow results

The comparison weighs retained earnings and avoided development against the business lost by waiting. When the gains outweigh those losses, pacing improves the lab’s finances even when customers already fund the race.

Gross profit often covers only a fraction of development.

The most useful comparable measure is gross profit divided by research and development expense. Gross profit is revenue after the company’s reported cost of sales; it is the amount available before research, sales, administration and other expenses. This ratio does not measure cash burn, but it reveals whether current sales are large enough to carry the current research organization.

OpenAI’s reported figures imply that gross profit covered about 29% of R&D in 2025. Development alone exceeded that contribution by $13.61B, before sales and administrative expenses. These are figures reported from private audited documents, rather than an audit or cash-flow statement made publicly available by OpenAI. R&D includes more than training. The reported operating loss and the much larger headline net loss should not be treated as interchangeable measures of cash consumed.

MiniMax’s H1 gross profit covered just 7% of research expense. Its filing attributes the increase in research expense primarily to cloud services for training and model iteration. Z.ai’s corresponding coverage was 11.8%. Both companies are growing, and both generate positive gross profit, but current sales leave most research expense to be funded from elsewhere. Their smaller scale makes them useful evidence about development economics, not direct estimates of the cost structure of OpenAI or Anthropic.

The SpaceX AI segment covered 66.8% of quarterly R&D with gross profit. It also reported $15.83B of quarterly capital spending. The segment combines Grok, X advertising, data licensing and cloud infrastructure: its revenue and gross margin cannot be assigned to Grok API requests. Its accounts explicitly include depreciation and share compensation in operating costs, so adding the entire R&D line to capital spending would count some infrastructure costs twice.

A low research-coverage ratio during a rapid expansion does not by itself establish that the investment will fail. That judgment requires future contribution, the useful life of the assets and the cost of obtaining capital. The financial table shows the development shortfall left after gross profit. Even profitable requests leave labs needing other funds when the development bill exceeds their contribution.

Growth shortens payback when total customer contribution rises.

OpenAI announced approximately $2B of monthly revenue and $122B of committed financing in March 2026. The financing figure is not a same-day cash balance. The announcement also described a business spanning consumer subscriptions, enterprise products and APIs, with infrastructure supplied through multiple partners. OpenAI ↗

Its reported annualized revenue exceeded $40B in August. A run rate extrapolates recent sales; it is not revenue already earned over a year. That newer scale anchors the illustrative scenarios, so the analysis does not assume the much smaller 2025 average business persists. Bloomberg, syndicated by Yahoo Finance ↗

Axios separately reported OpenAI’s $40B run rate. Madison Mills / Axios ↗

Anthropic announced a $47B revenue run rate in May and a $65B financing round. The latter included $15B of previously committed investments, so it should not be added indiscriminately to earlier funding announcements. These figures document rapid changes in the revenue base and access to funding; they do not reveal a model’s contribution margin. Anthropic ↗

Reuters subsequently relayed a Financial Times report that Anthropic expected positive adjusted operating income for a second consecutive quarter. Reuters relayed the report without independent verification. The separately cited margin above 80% was before distribution-partner sharing and training, and cannot be substituted for an all-in operating or cash margin. Even with those limits, the profitability report is material evidence against assuming persistent losses across all frontier labs. Sathvi G Bhat / Reuters, carried by MarketScreener ↗

Total revenue grows despite falling prices when additional paid usage more than offsets the price decline. Research investments support products and model variants beyond the flagship launch; the OpenRouter snapshot shows continued use of models below the benchmark frontier. A decline in the cheapest price available at a given intelligence level therefore does not imply the same decline in a lab’s revenue or profit.

Public financial statements do not identify the revenue with and without an individual model. When postponing a model sacrifices more new customer contribution than it preserves from existing models and saves in development, slowing destroys financial value. Avoiding repeated investment in a short-lived lead over close substitutes improves the trade in the other direction.

The financing bill depends on cash timing, not the largest headline number.

Revenue, gross profit, operating cash flow and capital spending answer different questions. Revenue records sales under accounting rules; the timing of cash collection is separate. Gross profit subtracts delivery costs but leaves development and overhead unpaid. Operating cash flow adjusts accounting profit for noncash items and working capital. Capital spending purchases assets used across model generations; depreciation spreads their accounting cost over time. Multi-year compute commitments describe future obligations and options, not a bill wholly paid at announcement.

A cash estimate therefore needs a bridge from operating income: add back noncash compensation and depreciation, account for collections, prepayments and supplier balances, then subtract the relevant cash investments and financing costs. Public private-lab reporting does not supply that bridge in sufficient detail for a reliable monthly forecast. A company-wide annual loss is not a defensible substitute.

Microsoft and Alphabet generated roughly $67B and $53B, respectively, after the cash capital spending shown above. Their broad software, advertising and cloud businesses provide funds for research whose return is captured outside a model API. These totals do not prove any particular AI investment is profitable; they show why a standalone-lab funding constraint cannot simply be imposed on a diversified parent. Microsoft ↗ · Alphabet / SEC ↗

Amazon’s trailing cash investment exceeded operating cash flow by about $7.6B. Meta was just above break-even after its quarterly capital spending and finance-lease principal. Positive operating cash flow leaves a financing gap when investment in capacity exceeds it. Periods and definitions differ in the table; the figures should not be ranked as a common annual “AI cash burn” measure. Amazon ↗ · Meta Platforms ↗

SpaceX’s consolidated H1 operating cash flow less gross cash property purchases was approximately negative $25B. That simple subtraction excludes rebates, capitalized interest and lease principal, and is not an AI-segment cash-flow figure. Its AI capital spending is visible separately, but segment operating cash flow is not. Without segment operating cash flow, public accounts do not establish Grok’s cash funding requirement.

CoreWeave’s Q2 adjusted EBITDA was $1.51B, yet depreciation and amortization were $1.39B and net interest expense $0.64B. Its operating cash flow was about $0.68B against $6.42B of cash property purchases. EBITDA excludes the cost of replacing and expanding the fleet and paying for its financing. Those costs leave a large funding requirement despite a high EBITDA margin. CoreWeave ↗

Microsoft’s FY2026 disclosure identifies $24.1B of commercial revenue involving OpenAI and a $6B receivable. That revenue includes revenue sharing, and the period differs from OpenAI’s 2025 accounts. It cannot all be labeled inference spending, nor does it represent cash already collected. A supplier receivable records the gap between recognized revenue and cash already collected. Microsoft / SEC ↗

At an industry level, the same economic expenditure appears in several places: a lab pays a cloud provider, which pays for chips, buildings and electricity. Adding lab compute spending to the supplier’s entire capital investment as independent costs would double count the chain. The relevant perspective must remain explicit: the lab’s cash obligation, the infrastructure owner’s investment, or the consolidated industry resource cost.

The final training run is only part of the development bill.

DeepSeek-V3’s technical report puts its official training run at 2.788M H800 GPU-hours, valued at $5.576M using an assumed $2 per GPU-hour. It explicitly excludes preceding research and ablation experiments. The figure describes a successful run’s compute consumption, not the cost of building the company, obtaining a fleet or discovering the method. DeepSeek-AI ↗

Meta reports 30.84M H100 GPU-hours for Llama 3.1 405B. Repricing those hours at the current $3.99 eight-GPU retail reference produces about $123M. This is a rental-equivalent calculation, not Meta’s actual historical payment: retail rental prices do not reproduce the cost structure of a large owned fleet or negotiated capacity. The hours themselves are a first-party disclosure. Meta ↗

Epoch’s historical research estimates put hardware and related equipment at 47–67% of selected models’ full development cost, staff at 29–49%, and energy at 2–6%. Its broader accounting includes experiments and researcher compensation. It also finds that retail-style cloud rental estimates substantially exceed amortized owned-hardware estimates in its sample. Those findings support separating training, development and fleet ownership; the study’s future cost projections are not observations of current spending. Ben Cottier et al. / Epoch AI ↗

A capacity calculation helps establish scale without pretending to know a proprietary training run. At an assumed $4 per GPU-hour, paying for 50,000 GPUs for 4,380 hours costs $876M; 100,000 costs $1.752B. This is a six-month capacity reservation whether the chips are efficiently used or idle. Research staff, data, experiments on other capacity and service operations add costs beyond this capacity calculation. Reusing the capacity across experiments spreads its cost across more work. These are scenarios, not disclosed lab invoices.

Mistral illustrates another business shape. Its spokesperson described new financing as supporting European computing capacity for both training and customer workloads, and gave a €1B revenue expectation for 2026. That expectation is not realized revenue, and comparable gross-profit or cash-flow accounts are unavailable in this evidence set. Its investment cannot be evaluated solely as an API business fighting for the highest general benchmark score. Le Monde ↗

The distinction matters for the $10B development block used below. It is a rounded six-month slice of a reported company-wide research budget, not a $10B training run. Smaller labs’ disclosed research programs operate at hundreds of millions, while the large lab and infrastructure portfolios reported above involve multiple billions or tens of billions. The bill to recover includes the whole research program: experiments, staff and capacity beyond the successful final training run.

The cost of an answer depends on how much paid capacity produces useful work.

For a concrete engineering estimate, NVIDIA reports Llama 3.1 70B in FP8 on four H100s producing 2,868.22 output tokens per second at concurrency 50, with 500 input and 2,000 output tokens per request. Time to first token is about 0.56 seconds and inter-token latency about 17 milliseconds. This is a particular older open-model workload, not a proxy for the intelligence or hidden serving stack of a current proprietary frontier model. NVIDIA ↗

Lambda lists a four-H100-SXM instance at $4.09 per GPU-hour, or $16.36 for the instance. The eight-GPU rate used for the training repricing above is $3.99. These are public retail references observed September 14, excluding tax, not a lab’s negotiated cloud bill. Lambda ↗

Rental cost per million output tokens = instance dollars per hour × 1,000,000 ÷ (3,600 × measured output tokens per second × realized capacity fraction)

The cost includes the measured prompt-processing work; input tokens are not silently treated as free. The denominator counts output tokens because that is how the benchmark reports system throughput. The additional realization factor represents paid fleet time that does not achieve that measured throughput, including slack capacity and uneven demand; it is not a replacement for the benchmark’s concurrency setting. NVIDIA’s metric definitions distinguish aggregate throughput from individual-user latency; optimizing the former does not guarantee the latter. NVIDIA ↗

At the central realization assumption, rental cost is about half a cent per request. For an explicitly hypothetical selling price of $3 per million input tokens and $15 per million output tokens, that request brings in 3.15 cents. At $0.20 input and $0.40 output, it brings in 0.09 cents. The first leaves a large rental contribution and the second does not cover this workload’s rental cost. These are pricing cases, not observed prices for this 70B open-model workload.

That remaining contribution must still cover service costs outside the rented instance, support, payment fees, free usage and development. A fleet operating at half the expected effective throughput doubles its rental cost per request. Reducing hardware cost or improving throughput expands the available contribution only to the extent the saving is not passed to customers through price competition.

DeepSeek supplies a different production example. For one day in February 2025 it reported enough occupied H800 capacity to imply $87,072 of rental-equivalent cost, serving 608B input and 168B output tokens. Pricing all that usage at R1 rates produced hypothetical revenue of $562,027, implying 84.5% contribution against that cost. The company explicitly said actual revenue was substantially lower because the mix included free access, cheaper V3 usage and discounts. Its “545%” figure used profit divided by cost, rather than the conventional profit divided by revenue. DeepSeek-AI ↗

The two examples should not be averaged into an industry margin. They differ in architecture, hardware, caching, batching, traffic and monetization. They show inexpensive serving under specific workloads and the overstatement produced by applying list prices to unbilled tokens. The most important unknown is often the realized mix of paid requests and the hardware needed to meet their latency requirements.

For an agentic task, a token price is only one part of the economic comparison. The full cost of a workflow includes every model call, reasoning token, tool invocation and required human review. If a workflow averages 1.5 calls per attempt and succeeds 80% of the time, its model cost per successful task is 1.875 times the single-call cost, before tools and review. That multiplier is an illustrative definition, not a measured failure rate. A higher-priced model earns its premium when it reduces the total resources required for a successful outcome.

Starting the same investment later lets customers cover more of it.

The earlier, controlled timing example uses $40B annual sales, a 40% contribution after serving costs, and two $10B development budgets. Revenue is anchored to a recent reported scale, the margin is an assumption informed by older accounts, and each upfront payment is an assumed simplification. These inputs deliberately remain separate from the observed financial table: no public source discloses this exact schedule.

The comparison holds sales, margins and total development spending constant. One path starts its second budget after 6 months; the other waits until month 12. There are no further budgets inside the 18-month horizon. This isolates financing timing; it does not give the slower lab equal capability as an empirical fact or assume an extra model would have no commercial value.

In the faster path, the first budget remains partly unrecovered when the next one begins. The problem is not the subtraction itself: the lab has to finance a new claim on its future earnings while an earlier claim is still outstanding. Its funding sources include investors, lenders, existing cash and other profitable businesses. These sources finance the next investment without establishing that the earlier model has paid for itself.

Real development spending occurs before and during launches, often under multi-year commitments. Moving payments to the beginning of a cycle makes the overlap visible, but exaggerates their lumpiness. R&D accounting includes noncash items such as stock compensation and depreciation. Without a cash-flow bridge and payment schedule, the correct description is a calibrated financing sensitivity, not an estimate of the exact amount a named lab must raise.

Six additional months produce $8B of contribution in this scenario. Uncovered spending at the second budget’s start falls from $12B to $4B. This is not an $8B reduction in peak capital needs. Both paths initially need $10B; peak uncovered spending is $12B versus $10B, a $2B difference. Nor is it an $8B profit gain: both paths have the same $4B residual before omitted costs at month 18. Waiting reduces peak financing pressure by giving customer contribution more time to cover the investment.

Margins, growth and budget size determine whether the mismatch persists.

At the central assumptions, one development budget requires 7.5 months of contribution to recover, before overhead or any later budget. A six-month cadence is faster than that recovery period. A 70% contribution margin would recover the first budget before the next cycle starts, while a 20% margin would stretch recovery to 15 months. The 20–70% range is a chosen sensitivity span, not a confidence interval for OpenAI.

The matrix shows the sensitivity to revenue assumptions: an old revenue figure overstates pressure on a growing lab, while an optimistic run rate understates it. Growth raises contribution only if delivery costs and other operating demands do not absorb it. Growing development budgets increase the amount that contribution must cover. An annualized sales figure is useful for scale, but treating it as guaranteed future cash would remove the very uncertainty the lab is financing.

Holding OpenAI’s reported 2025 expense structure and gross margin fixed, revenue would need to reach about $45B a year to cover R&D alone, or $62.1B to cover R&D plus the reported sales and administrative expenses. Those are algebraic break-even thresholds, not forecasts: they hold margin, compensation, research spending and revenue recognition fixed. They do not establish a cash runway.

Operating break-even revenue = (R&D + sales + administration) ÷ gross margin
Single-program recovery months = development budget ÷ monthly contribution

Existing cash extends the period a company finances the mismatch without new capital. It does not establish that customers have economically recovered the investment. Customer prepayments bring cash forward; supplier credit delays cash outflows; paying compensation in stock reduces current cash expense while diluting owners. Reusing assets retains productive value after one model loses its lead. These mechanisms are reasons to avoid inferring an imminent liquidity failure from an income statement alone.

The opposite mistake is to treat every unused GPU as instantly profitable inference capacity. Moving capacity depends on hardware suitability, software, deployment work and demand at acceptable prices. A percentage allocation by itself produces no revenue. The same applies to safety work: the financial consequence depends on the cash cost, infrastructure reuse and activity displaced. Allocating more compute to inference increases cash generation when paid demand covers the extra serving costs.

Pacing is attractive when it preserves earnings while deferring costly competition.

Underinvestment and overinvestment create different risks. Building too little puts customer retention at risk through inadequate capability or capacity, threatening the contribution expected to recover past spending. A lab that builds too much commits cash before demand is established and carries idle or rapidly depreciating assets. The next competitive deadline arrives before either uncertainty is necessarily resolved.

Coordinated pacing changes that decision by delaying competitors together. It extends the commercial life of existing capability while postponing the next development bill. Longer recovery time, later financing demands and less duplicative development are concrete financial advantages. Those benefits give labs a commercial reason to seek rules that restrain their competitors.

The cost of waiting includes the new demand, serving efficiencies and customer productivity gains forgone by delaying better models. Compliance and safety work require resources. Competitors outside the agreement remain free to win customers; idling contracted capacity does not cancel its payment obligations. A slower investment schedule is financially preferable only when its retained contribution, avoided spending and financing benefits outweigh those losses over a comparable horizon.

The quantities still needed for a reliable lab-specific forecast are cash margins by product, collections and supplier-payment schedules, the avoidable part of development spending, contractual capacity flexibility, and customer contribution with and without the next model. Those are material unknowns rather than small refinements. Until disclosed, ranges and explicit counterfactuals are more defensible than a precise model-by-model cash-flow curve.

The mix of motives behind an individual advocate’s position is unknown. This analysis examines financial incentives and does not estimate existential risk. Rules that prolong premiums and defer competitive spending improve returns on incumbents’ investments. Those stakes make the labs interested parties whose safety arguments must be tested independently.

Source inventory

External references are linked at their first use above. This inventory records their full identity and access limits; the downloadable input file retains every original URL and the figures used. Company announcements, private financial reporting, public interim accounts and engineering disclosures are kept distinct.

1. Ed Zitron / Where's Your Ed At. Exclusive: OpenAI Losses Increased Nearly 8X in 2025, With Spending Hitting $34 Billion. 2026-06-15. Full direct reporting read; underlying private audited statements not public. Linked FT corroboration paywalled.
2. OpenAI. OpenAI raises $122 billion to accelerate the next phase of AI. 2026-03-31. Public announcement, not audited statements
3. Bloomberg, syndicated by Yahoo Finance. OpenAI's Revenue Run Rate Tops $40 Billion Ahead of IPO. 2026-08-13. Indexed article lead available; full article not retrieved. The $40B figure is also reported in the accessible Axios August 17 article.
4. Madison Mills / Axios. Anthropic's revenue run rate reportedly surpasses $65 billion pre-IPO. 2026-08-17. Full article read; used only to cross-check this figure
5. Anthropic. Anthropic raises $65B in Series H funding at $965B post-money valuation. 2026-05-28. Public primary source
6. Sathvi G Bhat / Reuters, carried by MarketScreener. Anthropic tells investors it will be profitable for second straight quarter, FT reports. 2026-09-13. Full Reuters dispatch read; Reuters could not independently verify FT reporting; original FT article paywalled
7. MiniMax Group. Interim results for the six months ended 30 June 2026. 2026-08-26. Public 21-page announcement from investor relations; unaudited; full cash-flow statement absent
8. Z.AI Co., Ltd. / HKEX. Interim results for the six months ended 30 June 2026 (Chinese). 2026-08-31. Public unaudited announcement; Chinese labels translated
9. Space Exploration Technologies Corp. / SEC. Form 10-Q, quarter ended 30 June 2026. 2026-08. Public unaudited filing; AI segment includes X and cloud infrastructure
10. Microsoft. FY26 Q4 earnings press release and financial statements. 2026-07-29. Public primary source
11. Microsoft / SEC. Form 10-K, year ended 30 June 2026. 2026-07. Exact SEC-indexed disclosure read; full-page fetch failed
12. Alphabet / SEC. Second-quarter 2026 earnings release, Exhibit 99.1. 2026-07-22. Public primary source
13. Amazon. Amazon.com announces second-quarter results. 2026-07. Public primary source
14. Meta Platforms. Meta reports second-quarter 2026 results. 2026-07. Public primary source
15. CoreWeave. CoreWeave reports strong second-quarter 2026 results. 2026-08-11. Public primary source
16. DeepSeek-AI. DeepSeek-V3 Technical Report, version 2. 2025-02. Public primary source
17. DeepSeek-AI. Day 6: DeepSeek-V3/R1 Inference System Overview. 2025-03-01. First-party engineering disclosure, not audited financials
18. NVIDIA. NIM LLMs Benchmarking: Performance, Llama 3.1 70B, NIM 1.3.0. Undated benchmark table; accessed 2026-09-14. Public vendor benchmark; not reproduced locally
19. NVIDIA. NIM LLMs Benchmarking: Metrics. 2026-04-01. Public primary source
20. Lambda. GPU instances and public on-demand pricing. Accessed 2026-09-14. Public retail rates excluding tax, not negotiated lab prices
21. Meta. Llama 3.1 model card. 2024-07-23. Public primary source
22. Ben Cottier et al. / Epoch AI. How much does it cost to train frontier AI models?. 2024-06-03. Original research; historical estimates and forecasts, not current invoices
23. Le Monde. Mistral AI raises €3 billion in response to doubts over its strategic direction. 2026-09-08. Indexed excerpt with direct company comments; no comparable financial statements