Money owed by a company and money owed to it get mixed up. CoreWeave's $15.1B RPO and Cerebras's $24.6B RPO are customer backlog. Nscale owes ~$29B to suppliers (~178x its $163M revenue, single source) while customers owe it ~$103B.
Back to the trend of the day: The company behind the Claude AI asks to sell its shares to everyone
Poly Rating · from the trend of the day
AI Compute Obligations in S-1 Filings: Top 8 IPOs (2026) · 8 places
Anthropic has signed roughly $518 billion of compute deals, about 26 times the $19.7 billion it took in over the past year and more than three times the 7.4x that made Snap's 2017 cloud pledge look reckless.
- No. 1: Snowflake IPO (2020)
- 8 places
01Final list
The leading AIs built their own top-ten lists independently. Above is one combined result that is easy to check: every place has a reason, the AIs that included it and, where the AIs gave one, a source.
Order: from most to least AI agreement.
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
Snowflake IPO (2020)
$1.2B AWS minimum: 4.5x fiscal-2020 revenue of $264.7M, 3.0x the $402.7M trailing at pricing, 56% gross margin. A stricter reading of the filing finds only $169.5M across two agreements.
Included it: Claude, ChatGPT, Meta, Gemini, Z.ai, Kimi. Did not include: Grok, Nemotron.In their lists: Claude: 9 · ChatGPT: 3 · Meta: 5 · Gemini: 3 · Z.ai: 5 · Kimi: 6
What each AI noted 6
- Claudeplace 9
A cloud-native data company whose disclosed cloud minimums were only a few hundred million dollars and whose margin was rising to 62%.
Source ↗ - ChatGPTplace 3
The company generated $264.7 million in fiscal 2020 revenue and reported a 56% overall gross margin, a much thinner cushion than mature enterprise software typically enjoys.
Source ↗ - Metaplace 5
$1.0B cloud minimum over 5 years vs $264M FY19 revenue = 3.8x, 62% gross margin at listing
Source ↗ - Geminiplace 3
Bound itself to $1.2B in AWS infrastructure minimums against $402.7M in trailing revenue, requiring $350M/year peak cloud payments.
- Z.aiplace 5
The classic unicorn lien: $1.2 billion pledged to AWS over five years against a $264.8 million fiscal-2020 revenue book — 4.5x, at a 56% blended gross margin; it has since signed a $6 billion AWS deal in 2026.
Source ↗ - Kimiplace 6
$1.2B of minimum cloud purchases — 4.5x its final pre-IPO year revenue — the biggest cloud commitment anyone had signed at listing up to that point.
Source ↗
6 AIs put it here - 2
Snap IPO (2017)
$2B to Google Cloud plus $1B to AWS vs $404.5M of 2016 revenue: 7.4x (4.9x for Google alone), with a negative gross margin of about −12%. A circulating 19% figure misreads cost of revenue.
Included it: Claude, Gemini, Z.ai, Kimi. Did not include: ChatGPT, Meta, Grok, Nemotron.In their lists: Claude: 4 · Gemini: 2 · Z.ai: 3 · Kimi: 4
What each AI noted 4
- Claudeplace 4
A $2B five-year Google Cloud minimum against $404M of sales, with cost of revenue above revenue.
Source ↗ - Geminiplace 2
Pledged $3.0B across Google Cloud and AWS when generating only $404.5M in sales with an underwater gross margin of -12%.
- Z.aiplace 3
The original template: $3 billion promised to Google Cloud and AWS against $404.5 million of 2016 revenue — 7.4x, with only about 19 cents of each sales dollar left after infrastructure costs.
Source ↗ - Kimiplace 4
The original cloud pass-through: $3B locked into Google and AWS — 7.4x its 2016 revenue — with negative gross margins on day one.
Source ↗
4 AIs put it here - 3
Reddit IPO (2024)
$385M to AWS and Google Cloud through Sept 2026 vs $804M of 2023 revenue: 0.48x, 86% gross margin. Not to be confused with Google's $60M-a-year data deal, which is money coming in.
Included it: ChatGPT, Meta, Gemini, Z.ai. Did not include: Claude, Grok, Nemotron, Kimi.In their lists: ChatGPT: 8 · Meta: 4 · Gemini: 9 · Z.ai: 7
What each AI noted 4
- ChatGPTplace 8
Reddit's dependence on hosting infrastructure is economically different from buying long-lived accelerator capacity: variable traffic costs and fixed multi-year commitments are not interchangeable.
Source ↗ - Metaplace 4
$385M must-pay through Sept 2026 = 48% of $804M FY23 revenue, 86% gross margin
Source ↗ - Geminiplace 9
Locked in $385M in hosting commitments to AWS and GCP against $804M in revenue while maintaining an 86% software gross margin.
- Z.aiplace 7
Six months of revenue owed to the clouds: at least $385 million committed to AWS and Google Cloud through September 2026, on $804 million of 2023 sales.
Source ↗
4 AIs put it here - 4
Palantir direct listing (2020)
$1.49B six-year cloud-hosting minimum vs $743M of 2019 revenue: 2.0x, yet a 67% gross margin. Software economics with a heavy hosting bill.
Included it: ChatGPT, Meta, Gemini. Did not include: Claude, Grok, Nemotron, Z.ai, Kimi.In their lists: ChatGPT: 6 · Meta: 6 · Gemini: 8
What each AI noted 3
- ChatGPTplace 6
A software business with government and enterprise contracts is a useful counterweight to the AI infrastructure model, but its disclosed contractual obligations cannot automatically be treated as chip or cloud commitments.
Source ↗ - Metaplace 6
$1.49B six-year cloud host minimum = 2.0x $743M 2019 revenue, later reset to $5.6B ten-year
Source ↗ - Geminiplace 8
Disclosed $115M in non-cancellable cloud hosting obligations against $156.7M in revenue to support its enterprise AI platform.
Source ↗
3 AIs put it here - 5
Pinterest IPO (2019)
At least $750M to AWS through 2023 vs $755.9M of 2018 revenue: a full year of sales already spent, ~1.0x, with a 68% gross margin.
Included it: Claude, Gemini, Z.ai. Did not include: ChatGPT, Meta, Grok, Nemotron, Kimi.In their lists: Claude: 7 · Gemini: 6 · Z.ai: 6
What each AI noted 3
- Claudeplace 7
A $750M AWS minimum against $756M of revenue, about 1 times, yet a 68% gross margin.
Source ↗ - Geminiplace 6
Committed $750M to AWS over five years, matching 99% of its entire 2018 top line ($755.9M) on cloud visual search.
- Z.aiplace 6
A full year of sales, pre-spent: $750 million owed to AWS over six years against $755.9 million of 2018 revenue — and analysts flagged that cloud was eating roughly a fifth of every dollar.
Source ↗
3 AIs put it here - 6
Datadog IPO (2019)
Estimate, single source: a $225M AWS pledge vs $198M of 2018 revenue, ~0.8x, 76–78% margin. Disputed: another tally counts $1.4B of all commitments (7.1x), software included.
Included it: ChatGPT, Meta, Gemini. Did not include: Claude, Grok, Nemotron, Z.ai, Kimi.In their lists: ChatGPT: 5 · Meta: 7 · Gemini: 7
What each AI noted 3
- ChatGPTplace 5
A cloud-native monitoring platform is a useful control case: its S-1 reveals the cost structure of software delivered over third-party infrastructure rather than an AI lab committing hundreds of billions to accelerators.
Source ↗ - Metaplace 7
$1.4B non-cancellable commitments mainly cloud hosting vs $198M 2018 revenue = 7.1x, 78% gross margin
- Geminiplace 7
Pledged $225M in AWS cloud usage over three years while generating $198M in 2018 revenue with a 76% SaaS gross margin.
Source ↗
3 AIs put it here - 7
Lyft IPO (2019)
$300M AWS minimum through 2021 vs $2.16B of 2018 revenue: 0.14x, the lowest here. Its 42% gross margin (single source) reflects ride costs, not compute.
Included it: Claude, Gemini, Z.ai. Did not include: ChatGPT, Meta, Grok, Nemotron, Kimi.In their lists: Claude: 8 · Gemini: 10 · Z.ai: 9
What each AI noted 3
- Claudeplace 8
At least $300M of AWS spend through 2021, after an earlier $150M deal was doubled.
Source ↗ - Geminiplace 10
Bound itself to $300M in AWS compute fees against $2.16B in rideshare revenue with a compressed 42% gross margin.
- Z.aiplace 9
The smallest tab of the 2019 class: $300 million to AWS over three years — roughly 5% of each year's revenue, the closest thing this list has to renting compute cheaply.
Source ↗
3 AIs put it here - 8
Slack listing (2019)
At least $250M over five years to AWS vs $400.6M of revenue: 0.6x in total, $50M a year ≈12% of revenue. A normal software company. Single source.
Included it: Claude. Did not include: ChatGPT, Meta, Grok, Nemotron, Gemini, Z.ai, Kimi.In their lists: Claude: 10
What each AI noted 1
- Claudeplace 10
A $50M-a-year AWS minimum is only 12% of annual revenue, so Slack looks like a normal software company.
1 AIs put it hereno source given
02Worth knowing
What the AIs noticed while building the list.
About two-thirds of Anthropic's commitments cannot be cancelled. SpaceXAI can be cancelled on 90 days' notice, and Nscale and Lambda fees build up only as capacity is delivered. Revenue is pay-as-you-go while the payments are fixed.
The suppliers are also the owners. Google invests up to $40B while Anthropic commits $111.1B–$114.5B back to Google Cloud through 2033 and $110B to AWS through 2036.
Anthropic's $80.7B first-half loss is mostly non-cash. About $72B comes from fair-value marks on instruments tied to its rising valuation, and operating cash flow was positive at about $3.9B.
OpenAI is not ranked. Its draft S-1 was filed confidentially, and the figures quoted for it (about $600B by 2030, over $1.4T of infrastructure) come from the press, not a public prospectus.
03How we counted
In plain words, no formulas.
1The question
Top 10 Tech and AI Listings by Compute and Cloud Commitments vs. the Revenue to Pay for Them
2Why this topic
Commitments sit in prospectus footnotes and revenue sits in the income statement. Putting them side by side shows how much of each future sales dollar is already promised to AWS, Google Cloud and Nvidia.
3Where it applies
global software companies and frontier AI labs at the time of their IPO filing
4What we looked at
- Total commitments for compute, chips and cloud hosting disclosed in the prospectus notes
- Those commitments divided by revenue over the twelve months before filing
- Gross margin at filing, to show how much of each sales dollar is left after paying for compute
- SEC EDGAR (Anthropic S-1, Oct 2026, and earlier S-1s) — Purchase commitments, cloud minimums and leases in the notes and MD&A, plus revenue and cost of revenue. The sentence giving Anthropic's own tot
- Reuters — Adds up Anthropic's multi-year compute deals to at least $518B with six partners, and shows which of them cannot be cancelled
- Tom Tunguz — Adds the seven named compute contracts to $637B and separates out $265.6B of fixed or minimum guarantees. Derives gross margin from the cost-of-revenue note.
- Midgard Finance — Splits Anthropic's obligations by type: $269.4B of minimum cloud spend, $161.2B of Broadcom equipment leases and up to $206.4B that depends on capacity being delivered. Also gives a
- Mostly Metrics — Trailing twelve-month revenue of $19.7B, and an adjusted gross margin in the low 60s once free-tier compute and channel fees are moved back into cost of revenue. Also lease breakdowns
5Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Grok, Nemotron, Gemini, Z.ai, Kimi.
6How the places were set
The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.
04Each AI's own list
🇺🇸 Claude10 places · 4 notes
The same question, answered by Claude alone. The final list above combines all the AIs.
- 1Anthropic (Claude) IPO filing, 2026
About $637B of named compute contracts against $19.7B of trailing revenue, roughly 32 times, while H1 2026 gross margin jumped to about 85% from about 40% in 2025.
Source ↗ - 2OpenAI IPO (confidential draft S-1, 2026)
Press reports put compute plans at about $600B by 2030 and over $1.4T of infrastructure commitments, but the numbers are not yet in a public prospectus.
Source ↗ - 3CoreWeave IPO filing, 2025
$15.0B of signed but not-yet-started leases against $1.9B of revenue, about 8 times, funded by $7.9B of debt, with a 74% gross margin that hides heavy depreciation.
Source ↗ - 4Snap IPO filing, 2017
A $2B five-year Google Cloud minimum against $404M of sales, with cost of revenue above revenue.
Source ↗ - 5SpaceX (xAI) IPO filing, 2026
xAI lost $6.4B from operations on $3.2B of revenue in 2025, while group capex hit $20.7B and the same filing shows Anthropic paying $1.25B a month for Colossus capacity.
Source ↗ - 6Cerebras IPO filing, 2026 (refiled)
A chip maker that sells compute rather than renting it, with a 39% gross margin and a customer deal (OpenAI) reported at over $10B.
Source ↗ - 7Pinterest IPO filing, 2019
A $750M AWS minimum against $756M of revenue, about 1 times, yet a 68% gross margin.
Source ↗ - 8Lyft IPO filing, 2019
At least $300M of AWS spend through 2021, after an earlier $150M deal was doubled.
Source ↗ - 9Snowflake IPO filing, 2020
A cloud-native data company whose disclosed cloud minimums were only a few hundred million dollars and whose margin was rising to 62%.
Source ↗ - 10Slack IPO filing, 2019
A $50M-a-year AWS minimum is only 12% of annual revenue, so Slack looks like a normal software company.
What Claude noticed
Anthropic's gross margin is a derived number, not a reported one. Tunguz gets about 85% for H1 2026 against about 40% for FY2025, from the cost-of-revenue note, after the company stopped showing gross margin on the income statement.
The bill is large but partly flexible. Roughly two-thirds of Anthropic's commitments are non-cancelable. The rest can be cancelled on notice or accrue only as capacity is delivered. Revenue is consumption-based with little lock-in, while the payments are fixed.
One company's commitment is another's revenue. Anthropic's Lambda and Nscale deals total about $85B, and the SpaceX filing shows Anthropic paying $1.25B a month for Colossus. Suppliers, investors and customers overlap.
The same ratio looked very different in earlier IPOs. Pinterest's AWS minimum was about 1 times revenue and Snap's Google deal about 5 times. CoreWeave's leases reached about 8 times, and Anthropic's named contracts are about 32 times trailing revenue.
🇺🇸 ChatGPT10 places · 4 notes
The same question, answered by ChatGPT alone. The final list above combines all the AIs.
- 1Anthropic Claude AI IPO (2026)
At least $269.4 billion of minimum cloud spend plus $161.2 billion of equipment-lease payments: roughly 21.9 times the latest six-month-based trailing-twelve-month revenue when those two categories are combined.
Source ↗ - 2CoreWeave AI Cloud IPO (2025)
Its $15.1 billion of customer remaining performance obligations at the IPO period was about eight times annual revenue, but customer commitments are the revenue side of the infrastructure business, not its supplier bills.
Source ↗ - 3Snowflake Cloud Data Platform IPO (2020)
The company generated $264.7 million in fiscal 2020 revenue and reported a 56% overall gross margin, a much thinner cushion than mature enterprise software typically enjoys.
Source ↗ - 4Rubrik Cybersecurity IPO (2024)
A 77% overall gross margin in fiscal 2024 and $208.4 million in disclosed purchase obligations show a substantially lighter contractual burden than the frontier-AI examples.
Source ↗ - 5Datadog Cloud Monitoring IPO (2019)
A cloud-native monitoring platform is a useful control case: its S-1 reveals the cost structure of software delivered over third-party infrastructure rather than an AI lab committing hundreds of billions to accelerators.
Source ↗ - 6Palantir Technologies IPO (2020)
A software business with government and enterprise contracts is a useful counterweight to the AI infrastructure model, but its disclosed contractual obligations cannot automatically be treated as chip or cloud commitments.
Source ↗ - 7Cloudflare Internet Infrastructure IPO (2019)
Its business combines software with a distributed physical network, illustrating why hosting, bandwidth, equipment, and software costs should not be lumped together without checking the accounting definitions.
Source ↗ - 8Reddit Social Platform IPO (2024)
Reddit's dependence on hosting infrastructure is economically different from buying long-lived accelerator capacity: variable traffic costs and fixed multi-year commitments are not interchangeable.
Source ↗ - 9UiPath Automation Software IPO (2021)
Automation software is the high-margin comparison investors often have in mind when they value AI companies, but software licensing and frontier-model inference have very different underlying delivery costs.
Source ↗ - 10Arm Holdings Semiconductor IPO (2023)
Arm sells the intellectual property used in chips rather than operating a frontier AI assistant, making it an important contrast between monetizing compute demand and bearing the cost of serving it.
Source ↗
What ChatGPT noticed
Anthropic's disclosed supplier arrangements include $269.4 billion of minimum payments to Google, Amazon, and Microsoft and $161.2 billion of equipment-lease payments. The categories should be kept separate before adding them because they differ in contractual form and timing.
Anthropic's first-half 2026 compute and infrastructure expense was about 55% of revenue, but the reported customer-serving gross margin was approximately 85%. These figures answer different questions: the first captures a broader infrastructure expense category, while the second focuses on customer-serving economics.
CoreWeave's $15.1 billion of remaining performance obligations at the end of 2024 grew to $60.7 billion by the end of 2025. That is a customer-demand backlog, not a direct measure of the company's own future chip and data-center bills.
Snowflake's overall gross margin was 56% in fiscal 2020, compared with 77% for Rubrik in fiscal 2024. Different business models and accounting periods prevent interpreting that 21-percentage-point gap as a pure measure of software quality.
🇺🇸 Meta10 places · 3 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1Anthropic 2026 IPO (ANT)
$518B locked-in = 27x its $19B TTM revenue – $114.5B Google + $110B AWS + $161B Broadcom alone
Source ↗ - 2OpenAI 2026 Confidential Filing
$665B footnote vs $50B run-rate = 13.3x sales, gross margin only 33% in 2025
- 3CoreWeave 2025 IPO (CRWV)
$80B lease commitments + $35.6B debt vs $1.9B revenue = 42x leverage, true margin 30% not 74%
- 4Reddit 2024 IPO (RDDT)
$385M must-pay through Sept 2026 = 48% of $804M FY23 revenue, 86% gross margin
Source ↗ - 5Snowflake 2020 IPO (SNOW)
$1.0B cloud minimum over 5 years vs $264M FY19 revenue = 3.8x, 62% gross margin at listing
Source ↗ - 6Palantir 2020 Direct Listing (PLTR)
$1.49B six-year cloud host minimum = 2.0x $743M 2019 revenue, later reset to $5.6B ten-year
Source ↗ - 7Datadog 2019 IPO (DDOG)
$1.4B non-cancellable commitments mainly cloud hosting vs $198M 2018 revenue = 7.1x, 78% gross margin
- 8C3.ai 2020 IPO (AI)
$403M partner commitment disclosed vs ∼$157M FY20 revenue = 2.6x, heavy AWS/Azure/GCP pass-through
- 9Confluent 2021 IPO (CFLT)
Cloud infra purchase obligations primary cash need – $200M+ yearly host burn at filing, 66% gross margin
- 10Nebius 2024 IPO via SPAC (NBIS)
$40B+ long-term customer commitments secured vs $510M 2025 revenue = inverse leverage – neocloud sells commitments
What Meta noticed
Anthropic's 85% gross margin is by function – PYMNTS notes cost of revenue $2.38B on $16.3B. PitchBook analysis says true compute cost was $0.71 per $1 revenue in Q1 ($0.56 in Q2), implying 29% and 44% margins when training compute is included – the most important trend is the 15-cent quarterly improvement.
Circular financing: Google commits $40B ($10B cash + $30B milestones) then Anthropic commits $111.1B-$114.5B back to Google Cloud through 2033, AWS $110B through 2036, plus 16% marketplace toll on revenue flowing through Bedrock/Vertex.
Off-balance-sheet tracker: $3.56T total across Alphabet, Meta, Microsoft, Nvidia, Oracle, Amazon, Broadcom as of Oct 22 2026 – $2.03T purchase commitments + $1.25T leases not yet commenced – dwarfs AI lab revenue.
🇺🇸 Gemini10 places
The same question, answered by Gemini alone. The final list above combines all the AIs.
- 1Anthropic (2026, US)
Committed $265.6B in guaranteed compute minimums against $16.3B in recent revenue—a record 16.3x ratio turning future sales into hyperscaler pass-throughs.
- 2Snap Inc. (2017, US)
Pledged $3.0B across Google Cloud and AWS when generating only $404.5M in sales with an underwater gross margin of -12%.
- 3Snowflake (2020, US)
Bound itself to $1.2B in AWS infrastructure minimums against $402.7M in trailing revenue, requiring $350M/year peak cloud payments.
- 4Palantir Technologies (2020, US)
Carried $1.5B in multi-year third-party cloud hosting obligations compared to $742.6M in prior-year revenue.
- 5SentinelOne (2021, US)
Incurred $95M in cloud hosting commitments alongside $93.1M in revenue, balancing security compute against a 58% gross margin.
Source ↗ - 6Pinterest (2019, US)
Committed $750M to AWS over five years, matching 99% of its entire 2018 top line ($755.9M) on cloud visual search.
- 7Datadog (2019, US)
Pledged $225M in AWS cloud usage over three years while generating $198M in 2018 revenue with a 76% SaaS gross margin.
Source ↗ - 8C3.ai (2020, US)
Disclosed $115M in non-cancellable cloud hosting obligations against $156.7M in revenue to support its enterprise AI platform.
Source ↗ - 9Reddit (2024, US)
Locked in $385M in hosting commitments to AWS and GCP against $804M in revenue while maintaining an 86% software gross margin.
- 10Lyft (2019, US)
Bound itself to $300M in AWS compute fees against $2.16B in rideshare revenue with a compressed 42% gross margin.
🇨🇳 Z.ai10 places · 4 notes
The same question, answered by Z.ai alone. The final list above combines all the AIs.
- 1OpenAI (S-1 filed confidentially May 22, 2026)
Roughly $1.4 trillion of compute promises against $13 billion of 2025 revenue — about 108 times the money it actually took in, and its S-1 is still sealed.
Source ↗ - 2Anthropic IPO S-1 (Nasdaq: ANT, filed public Oct 26, 2026)
At least $518 billion of 'cannot-be-canceled' compute deals — about 32 times the $16.3 billion it booked in the first half of 2026 — with Amazon and Google on both sides of the table.
Source ↗ - 3Snap IPO (2017)
The original template: $3 billion promised to Google Cloud and AWS against $404.5 million of 2016 revenue — 7.4x, with only about 19 cents of each sales dollar left after infrastructure costs.
Source ↗ - 4CoreWeave IPO (2025)
The middleman's ledger: $15.1 billion of contracted obligations on its books against $1.9 billion of 2024 revenue at filing — and roughly $50 billion of debt within 18 months of listing.
Source ↗ - 5Snowflake IPO (2020)
The classic unicorn lien: $1.2 billion pledged to AWS over five years against a $264.8 million fiscal-2020 revenue book — 4.5x, at a 56% blended gross margin; it has since signed a $6 billion AWS deal in 2026.
Source ↗ - 6Pinterest IPO (2019)
A full year of sales, pre-spent: $750 million owed to AWS over six years against $755.9 million of 2018 revenue — and analysts flagged that cloud was eating roughly a fifth of every dollar.
Source ↗ - 7Reddit IPO (2024)
Six months of revenue owed to the clouds: at least $385 million committed to AWS and Google Cloud through September 2026, on $804 million of 2023 sales.
Source ↗ - 8Dropbox IPO (2018)
The $500 million AWS lien that taught Dropbox to build its own hardware: about 45% of a year's revenue committed for five years — then a slow, deliberate exit to in-house infrastructure.
Source ↗ - 9Lyft IPO (2019)
The smallest tab of the 2019 class: $300 million to AWS over three years — roughly 5% of each year's revenue, the closest thing this list has to renting compute cheaply.
Source ↗ - 10Figma IPO (2025)
The anti-Anthropic: a '$300,000-a-day' AWS bill that made headlines, but it is really just 12% of revenue wrapped in a 91% gross margin — software economics, not compute pass-through.
Source ↗
What Z.ai noticed
The scale moved two orders of magnitude in one IPO season: the 2017–2020 class pledged roughly 0.1x–7x revenue to the clouds, while Anthropic's 2026 prospectus discloses at least $518 billion — about 32x one half-year of revenue — and Reuters found most of those deals 'cannot be canceled.'
The suppliers own the customer: Amazon holds about 21% of Anthropic and Google about 15% while selling it the compute, and one analyst case study tallies roughly $25 billion of hyperscaler investment returning an estimated $75–105 billion in infrastructure fees — a negative effective cost basis.
The margin story inverted since Snap: in 2017 Snap filed with 7.4x commitments at a 19% gross margin; Anthropic filed claiming >80% gross margin — but that figure is calculated before revenue shared with distribution partners and before model-training costs, which is precisely what the commitments buy.
The same dollars appear on both sides of the market: Cerebras listed at ~48x trailing revenue with $24.6 billion of performance obligations (including >$20B from OpenAI for 750MW through 2028), Lambda's backlog jumped to $50 billion on Anthropic's $35 billion deal, and Google will pay SpaceX $920 million a month for GPU access — the sellers are IPO-ing on the buyers' obligations.
🇨🇳 Kimi10 places · 4 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1Nscale (2026 S-1, NYSE: NSCL)
$29B of equipment, construction and colocation commitments — roughly 178x the $163M it earned in the year before filing — with only 25,000 of 461,000 contracted GPUs switched on.
Source ↗ - 2Anthropic (2026 S-1, Nasdaq: ANT)
Signed away an estimated $518–637B of future compute — 26–32x trailing revenue — while its $80.7B first-half loss is mostly mark-to-market marks, not cash burn.
Source ↗ - 3CoreWeave (2025 S-1, Nasdaq: CRWV)
The template: ~$15B of not-yet-started GPU and data-center leases — about 8x 2024 revenue — while still keeping 74 cents of each revenue dollar.
Source ↗ - 4Snap (2017 S-1, NYSE: SNAP)
The original cloud pass-through: $3B locked into Google and AWS — 7.4x its 2016 revenue — with negative gross margins on day one.
Source ↗ - 5Nebius Group (2024 relisting, Nasdaq: NBIS)
Capex ran at 883% of revenue the year it relisted — the plan, not yet the contract, so it ranks below those with filed obligations.
Source ↗ - 6Snowflake (2020 S-1, NYSE: SNOW)
$1.2B of minimum cloud purchases — 4.5x its final pre-IPO year revenue — the biggest cloud commitment anyone had signed at listing up to that point.
Source ↗ - 7GlobalFoundries (2021 F-1, Nasdaq: GFS)
$4.7B of unconditional purchase commitments — essentially 1x its entire prior-year revenue — the chip-foundry version of the same bet.
Source ↗ - 8Cerebras (2026 S-1, Nasdaq: CBRS)
Spent $383M building cloud capacity in 2025 against $510M of revenue — and its S-1 never totals what it owes AWS and its landlords.
Source ↗ - 9Astera Labs (2024 S-1, Nasdaq: ALAB)
Filed with essentially nothing owed: $27.5M of purchase commitments by end-2024 — 0.24x revenue — because the foundry carries the capex.
Source ↗ - 10Figma (2025 S-1, NYSE: FIG)
$41.7M of cloud-hosting commitments — 0.06x revenue — the ranking's control group: software that keeps ~90 cents of each dollar.
Source ↗
What Kimi noticed
The AI listings make Snap 2017 look prudent: Snap's $3B cloud commitment was 7.4x revenue with negative margins; Nscale 2026 is ~178x and Anthropic 2026 is ~26–32x, yet both found eager IPO demand because AI scarcity repriced the landlord model Snap was punished for.
Anthropic's headline $80B-plus first-half loss is mostly accounting: roughly $72B is mark-to-market on its own rising share price (Bartz settlement, charitable shares, related-party marks), and operating cash flow was positive — a twist that separates the GAAP panic from the cash reality.
Customer-side and supplier-side commitments are routinely mixed up: Cerebras's $24.6B RPO and Nscale's $103.4B TCV are money owed TO the company, while this ranking measures money owed BY it — Nscale owes ~$29B to Dell, landlords and lenders while being owed ~$103B by Microsoft and Anthropic.
Margin at filing is the tell for who keeps the dollar: CoreWeave kept 74% gross margin as a GPU landlord, Anthropic ~40% as a model lab, Nscale was negative, and Figma ~90% — the same AI wave produced all four models in a single 18-month IPO window (2025–2026).