Wholestory

The first frontier-lab IPO has moved back about five weeks. Anthropic had been expected to make its prospectus public in the week of 7 September; Reuters reported on 4 September, citing people familiar, that this is now expected late September, with marketing from mid-October and the listing days before the November midterms. A bank credit facility must be finalised first. Anthropic declined to comment. Until a prospectus exists its revenue figures are annualised run rates shown to investors, not audited ones — and a full-text search of SEC records on 7 September still returns no Anthropic registration statement.

The chip supplier's exposure to its own customers now has a filed number. Broadcom's quarterly report for the period ended 3 May, filed on 9 June, discloses that it arranged for an investor partner to buy AI racks built on Broadcom's accelerators and lease them to a customer, and that Broadcom then entered a backstop agreement covering that customer's five-year lease obligations, "with a maximum exposure of $29 billion". The backstop grows as racks are deployed and shrinks as the tenant pays; on default Broadcom can assume the lease or force a sale of the racks. On the 2 September earnings call an analyst put the same figure to the company. Finance chief Amie Thuener said those numbers "remain true" and refused to give any cap for future arrangements: "I can't give you an overarching look at what's the max."

Broadcom reported net revenue of $29.59 billion for the quarter to 2 August, up 86%, and said $16.7 billion of it was AI semiconductor revenue, up 221%. Note what is filed and what is not. The audited segment table shows two lines — semiconductor solutions $20.84 billion, infrastructure software $8.75 billion — and the AI figure is a management carve-out inside the first, disclosed in the chief executive's quote and reconciled to nothing. Broadcom's own capital spending was $532 million: the capital intensity sits with its customers. Fourth-quarter guidance of $34.8 billion came in under the $35.03 billion expected.

The price side of the bubble question moved. Silicon Data's LLM Token Expenditure Index — a daily, usage-weighted benchmark for what a million language-model tokens actually cost, blending posted provider prices with consumption observed across routing gateways — fell to 97 cents on 31 August, its lowest since the index began late in 2025 and more than half below its summer high. Cheap open-weight rivals, chiefly Moonshot's Kimi K3, and OpenAI's late-July price cuts are the named drivers. For labs whose compute commitments are fixed for years, a falling output price compresses revenue without touching cost. It had recovered to 99 cents by 5 September.

Anthropic is heading toward its IPO citing an annualized revenue run rate above $65 billion; OpenAI puts its own above $40 billion. Most of the $25-billion gap, The Information and accounting professor Francine McKenna find, is bookkeeping rather than business. When a customer buys Claude through a cloud partner, Anthropic treats itself as the principal and records the full sale as revenue, expensing the partner's cut; OpenAI treats itself as the agent and books only its net share. Neither is wrong. But even restated to a net basis — a roughly 6-to-10-percent haircut — Anthropic would still sit $19-21 billion ahead. How it justifies the gross treatment should surface in its prospectus.

OpenAI said it had confidentially submitted a draft registration statement on Form S-1 to the SEC, under Rule 135 — a week after Anthropic's own confidential filing, and like it not a public prospectus: no financials, share count, price range or timetable. OpenAI added it had not settled on timing and that a listing 'may be a while,' since some plans are easier to pursue as a private company. It followed the company's March round of $122 billion in committed capital at an $852 billion post-money valuation, anchored by Amazon, Nvidia and SoftBank with Microsoft. Both frontier labs have now opened the IPO process without yet exposing audited accounts.

The European Central Bank, analysing why American hyperscalers have arrived in Europe's bond market, put the spending in a frame this page has not had from an institutional source: capital expenditure projected to exceed $1 trillion in total by 2028, equal to about 3 percent of current annual US GDP. Its accompanying observation is the one that explains the borrowing — planned investment has grown too large to finance out of internally generated cash flow, so the companies are shifting from self-funding toward external finance. The trillion is a projection rather than a measurement, and the ECB is citing it rather than producing it; what belongs to the ECB is the comparison and the reading of what it does to corporate balance sheets.

Nvidia's second quarter to 26 July brought $96.2 billion of revenue, up 18 percent on the previous quarter and 106 percent on the year, at a 75 percent gross margin — $89.0 billion of it from data centres, up 117 percent year on year. GAAP net income was $59.7 billion and diluted earnings $2.46 a share. The company returned about $26 billion to shareholders in the quarter and has $99 billion of buyback authorisation left. The guide for the current quarter is $108 billion, plus or minus two percent, and it assumes no data-centre compute revenue from China at all — a market Nvidia has now written out of its own forecast. Jensen Huang's framing of the quarter was that “compute is revenue”, and that where one lab drove the buildout a year ago, multiple frontier labs are now scaling in parallel.

Alongside its results, Nvidia said it had agreed partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to set up independent compute-infrastructure financing platforms, intended to mobilise more than $500 billion of third-party capital for the AI buildout — “over time, subject to definitive agreements”, which is to say nothing is signed and no capital has moved. The structure is the story rather than the number: the company selling the accelerators is now organising the outside capital that will pay for them, one step further along the same axis as its equity stakes in customers.

Anthropic will tell IPO investors its potential revenue opportunity exceeds $30 trillion, the Wall Street Journal reported on 25 August, citing people familiar — larger than the $28.5 trillion SpaceX showed before its own offering. The company more than doubled revenue to $11.6 billion in the second quarter, and could seek as much as $100 billion at about a $2 trillion valuation, against SpaceX's $86 billion and $1.77 trillion debut in June. Nothing is filed and no figure is final. NYU's Aswath Damodaran called SpaceX's comparable estimate “reaching the end of what's plausible”; the stock trades below its offer price.

Regulators have gone looking at the leverage behind the AI trade. The SEC has subpoenaed Bank of America, Citi, Goldman Sachs and JPMorgan Chase over Situational Awareness, the fund that lost about 67 percent of its portfolio in July, asking for the timing of its trades and its communications with lenders about borrowed money, and warning the banks to preserve records. The mechanism was concentration: $5.7 billion of SanDisk and $5.6 billion of Micron were more than half its US book at the end of June, and falls of 47 and 29 percent in July forced margin calls and a sale of the whole public book to Citadel at roughly a 10 percent discount. The inquiry is early, the fund is accused of nothing and says it will cooperate. Note what is being investigated: not AI demand, but what was borrowed against it and what lenders were told.

Anthropic surpassed OpenAI in quarterly revenue for the first time. In the three months to June it reported about $11.6 billion, more than double its roughly $4.73 billion first-quarter figure; OpenAI reported $6.7 billion, up 18 percent from $5.7 billion — growth that disappointed investors hoping it would close the gap. The bottom lines diverged further: OpenAI's operating loss widened to $12.3 billion from $9.3 billion, while Anthropic recorded a small adjusted operating profit of about $559 million — though it excludes stock-based compensation and has not said how it calculates the figure, and all the numbers are preliminary and press-reported, not audited. The shift, driven by Claude Code's traction with corporate customers against decelerating ChatGPT growth, has pushed OpenAI into a reactive posture, with management departures including chief revenue officer Denise Dresser after less than a year, following the earlier exits of operating chief Brad Lightcap and Fidji Simo.

Nvidia disclosed in a Form 8-K that on 17 August it entered into multiple residual value guaranties with SB Energy covering leases for approximately 4.25 gigawatts of IT load at the PORTS-Pike campus in Pike County, Ohio, its aggregate payment obligation “cumulatively capped at $105 billion.” OpenAI is the tenant on a 20-year lease of a site SB Energy will build, own and operate, with the first capacity expected from 2028. The figure carries the whole arc of the deal: the Wall Street Journal reported in late July that Nvidia was in talks to guarantee roughly $250 billion; on 14 August it reported the plan cut to less than $120 billion after investors objected to how much of Nvidia's own balance sheet was standing behind demand for its chips; three days later the executed instrument capped at $105 billion and covered only the first phase, with an option, at Nvidia's sole discretion, to add credit support for a further 3.8 gigawatts. The filing states precisely what triggers payment: OpenAI's insolvency resulting in a default under a lease, or OpenAI's failure to make lease payments. On either, Nvidia pays the shortfall between the lease's guaranteed minimum value and whatever is recovered through a replacement lease or a sale, and may assume the lease, require the lessor to relet, start a sale process, allow termination, or defer those remedies for up to a year while paying project costs. The obligations end at the earliest of the twentieth anniversary of each lease, OpenAI terminating it, or OpenAI achieving a satisfactory credit rating — the absence of which is what makes the guarantee necessary in the first place. OpenAI has agreed to reimburse and indemnify Nvidia for any amount it actually pays. Nvidia filed the arrangement under Item 2.03, the item covering an off-balance-sheet obligation, and will file the form of the agreements with its quarterly report for the period ended 26 July. Separately, Nvidia is investing $1.5 billion in SB Energy alongside SoftBank and OpenAI. Goldman Sachs and JPMorgan advised SB Energy; Morgan Stanley advised Nvidia.

Situational Awareness, the fund founded in July 2024 by former OpenAI researcher Leopold Aschenbrenner and named after his widely read essay, was forced by margin calls to sell its entire portfolio of listed positions to Ken Griffin's Citadel at a discount. Its assets fell from about $45 billion at the start of July to around $10 billion. The losses came from both sides of the book: long positions in AI infrastructure such as SK Hynix and CoreWeave fell while short positions in software companies including Adobe moved sharply against it, on reported leverage of as much as 400%. Bank of America, Goldman Sachs and JPMorgan Chase, its prime brokers, worked to reduce the positions. About two-thirds of the fund's holdings were listed long and short positions; the rest were private stakes dominated by a multibillion-dollar investment in Anthropic, which the firm said it was not marketing. The counterweight belongs in the record too: the episode has so far stayed contained — Citadel absorbed the book, the named stocks have since rallied, and Ninety One's Sahil Mahtani attributes the collapse to the fund's own risk management rather than to systemic leverage, arguing that elevated earnings expectations, not borrowing, are the AI trade's main risk to markets.

Nvidia signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish what it called the first compute-financing platforms of their kind at global scale, aiming to mobilise more than $500 billion of third-party capital to fund the buildout across its ecosystem — frontier labs, enterprises and AI clouds. The design creates dedicated pools of long-term capital that lend to Nvidia's own customers, treating GPUs the way markets treat toll roads or real estate: a bankable, income-producing asset class rather than rapidly depreciating hardware. The arrangement is the largest expression yet of the 'circular financing' structure already in the record — a supplier orchestrating the capital that buys its own product, extending the pattern of Nvidia's up-to-$100 billion commitment to OpenAI. Chief executive Jensen Huang framed it as a milestone in which 'compute is revenue' and Nvidia hardware is 'continuously improved through CUDA software, extending its useful life' — the announcing party's promotional case, and a direct counter to the depreciation criticism (Michael Burry's useful-life claim) already tracked here. The agreements 'remain subject to execution of the final agreements': these are MOUs and an intention to build capital pools over time, not closed financings. The $500 billion figure and the six-firm roster are corroborated beyond Nvidia's own release by Reuters, the Wall Street Journal and the Financial Times.

In its first earnings report since its June initial public offering, SpaceX (SPCX) reported second-quarter 2026 revenue of $7.8 billion, up 92% year over year and above the roughly $6.8 billion consensus, against a net loss of $541 million. Capital expenditure soared more than sixfold to $18.4 billion — more than double quarterly sales and far above the roughly $13 billion analysts had forecast — with nearly $16 billion, well over 80%, flowing to the xAI/AI-compute business that merged into SpaceX in February 2026. The AI segment (SpaceXAI) posted quarterly revenue of $2.56 billion, up 247% year over year, and an operating loss of $1.26 billion, while Starlink revenue grew 66%. The shares fell about 8% after the report, leaving them more than 20% below their June 12 first-trade price. The $18.4 billion capex, the segment figures and a $354 million Memphis litigation accrual are attributed by CNBC and CNN to SpaceX's Q2 10-Q, which the harness could not retrieve directly (SEC 403; no Wayback snapshot); analysts expect full-year capital spending above $45 billion.

AMD reported record second-quarter 2026 revenue of $11.5 billion, up 50% year over year, led by data-centre revenue that more than doubled — up 107% — to $6.7 billion, a reading of AI demand across the chip layer alongside Nvidia. Client revenue was $3.8 billion, up 6%, while gaming revenue fell 31% to $779 million, which chief executive Lisa Su attributed to the late stage of the console cycle and higher component and memory costs.

Palantir reported second-quarter 2026 revenue of $1.935 billion, up 93% year over year, with U.S. commercial revenue up 149%. GAAP net income was $1.062 billion, a 55% margin, and adjusted free cash flow $1.220 billion, a 63% margin; the company raised full-year 2026 revenue guidance to $8.150–8.158 billion. Closed total contract value reached $3.373 billion (up 49%), U.S. commercial contract value a record $2.132 billion (up 153%), and cash and short-term Treasuries stood at $9.2 billion. These are filed figures from the company's own earnings release, and they stand as this page's cleanest example of AI-linked revenue that is both large and GAAP-profitable — a counter-case to the frontier labs' deepening losses.

Amazon posted Q2 2026 net sales of $200.6 billion and net income of $62.6 billion, which includes a pre-tax income of $53.4 billion primarily from its investments in Anthropic. AWS generated $42.2 billion in sales, expanding 37% year-over-year. CEO Andy Jassy noted that the AI and homegrown chips units each exceeded a $25 billion annual revenue run rate. However, Q2 capex reached $54.2 billion, pushing trailing-twelve-month free cash flow to negative $7.6 billion. The company raised its 2026 capital expenditures forecast to $220 billion, up from the $200 billion projected in February.

The same filing recorded net income of $112.193 billion and diluted earnings of $9.11 a share — the largest quarterly profit ever reported by a company. Of it, $99.031 billion is a net gain on equity securities, a line sitting below operating income and consisting largely of unrealized revaluation of private holdings; the comparable figure a year earlier was $1.286 billion. Alphabet's operating income was $40.770 billion. Excluding the gain, adjusted earnings were $2.85 a share, short of the $2.89 consensus. CNBC and Fortune report the markup came mainly from Alphabet's stakes in Anthropic, whose private valuation reached $965 billion in May, and in SpaceX; Alphabet does not disclose how much came from each. Two disclosures landed in one report: the revaluation of holdings that include Anthropic, and an 82% rise in Google Cloud revenue in a segment whose customers include Anthropic, which has committed to buying at least five gigawatts of Google Cloud capacity. The balance sheet records the revaluation separately — non-marketable securities rose from $68.687 billion to $131.461 billion in six months, while long-term debt rose from $46.547 billion to $98.165 billion over the same period. The quarterly report filed a day later narrows the picture without closing it. It puts the carrying value of Alphabet's non-marketable holdings in private companies at $124.3 billion, and discloses separately a stake in SpaceX — public since its listing — worth $94.1 billion, of which $80 billion is subject to early-release sale restrictions and $14.1 billion stays restricted until the third quarter of 2027. Yahoo Finance later clarified the $99 billion mark-to-market gain was primarily the result of SpaceX's June 2026 IPO at a $1.77 trillion valuation (Alphabet holding a 4.9% stake valued at $94.0 billion) and a surge in Anthropic's private-market valuation to $965 billion.

Microsoft closed its fiscal year with revenue of $90.007 billion for the quarter, up 18%, operating income of $40.603 billion, up 18%, and net income of $35.766 billion — and unlike Alphabet a week earlier it funded a record quarter of building out of its own cash. Capital expenditure including finance leases was $41 billion, up about 69% year over year and taking the fiscal year to $145.3 billion, with roughly two thirds going to short-lived assets, mainly CPUs and GPUs. Operating cash flow rose 30% to $55.4 billion; free cash flow fell 23% to $19.64 billion but stayed comfortably positive, and Microsoft still returned $10.2 billion to shareholders in the quarter. The return side is the part that moved. Azure passed $100 billion of revenue for the fiscal year for the first time and grew 43% in the quarter, accelerating from 40%; commercial remaining performance obligation reached $678 billion, up 84%, and — answering the concentration question directly — up 25% excluding OpenAI, with a weighted-average duration of 2.3 years and about 30% due to be recognised within twelve months. None of that is an AI revenue line: Microsoft discloses Azure only as a growth rate, Microsoft Cloud as a company-defined aggregate of $59.3 billion, and Intelligent Cloud as a $39.306 billion segment that is not AI-only. Its separate disclosures on OpenAI are firmer — $24.1 billion earned from the partnership through the fiscal year including revenue-sharing payments, $6 billion owed by OpenAI, $13 billion of funding committed and $11.9 billion provided. Shares rose about 8% after hours and were up roughly 14% at the next open, having gone into the report down about 19% for the year. Seven days after a company that beat on revenue was sold for planning to spend more, a company that beat while holding its spending flat was bought.

Effective with its 2027 fiscal year, Microsoft is extending the estimated useful lives of its data centres and office buildings from 15 years to 25 — the first depreciation-schedule change recorded on this page, and the exact manoeuvre Michael Burry accused the industry of in November 2025. Chief financial officer Amy Hood set it out on the earnings call: the extension reflects 'our operating history and expected use of these assets', 'affects only the timing of future depreciation and is expected to have a minimal benefit to FY27 operating income'. The larger consequence is not to earnings but to the headline number the market watches. Because more future data-centre leases will now be classified as operating leases rather than finance leases, and finance leases count as capital expenditure while operating leases do not, Microsoft's stated calendar-2026 expectation falls to about $175 billion from the roughly $190 billion guided in April — while, in Hood's words, 'outside of this useful life impact, our calendar year 2026 CapEx investment expectations remain unchanged'. Nothing about what is being built has changed; about $15 billion of it is now counted somewhere else. Two things keep the accusation and the change apart. The reclassification does not flatter free cash flow, because operating-lease payments come out of operating cash flow instead. And the assets whose lives were extended are buildings, not the chips inside them: roughly two thirds of Microsoft's capital spending this quarter went to CPUs and GPUs, which are depreciated on far shorter schedules and were untouched. Meta's results, published the same day, contain no change to any useful-life assumption and record depreciation up 46% year over year.

Meta's second quarter is the cleanest illustration yet of what the buildout costs a company that has no cloud business to sell it back through. Revenue rose 28% to $60.801 billion — a record — while operating income fell 8% to $18.775 billion and net income fell 14% to $15.848 billion, taking the operating margin to 31% from 43%. Earnings of $6.18 a share missed consensus of between $7.14 and $7.22. Free cash flow fell to $784 million from $8.55 billion a year earlier, a fall of about 91%, and the whole of it is capital spending: operating cash flow actually rose, to $31.862 billion from $25.561 billion, while capital expenditure including finance-lease principal reached $31.08 billion against $16.538 billion of property and equipment purchases a year before. Depreciation and amortisation rose 46% to $6.356 billion and total costs rose 55%, including $2.40 billion of legal charges and $1.18 billion of severance from the May headcount reduction; excluding those, chief financial officer Susan Li said operating income would have risen 9%. Meta narrowed its 2026 capital-expenditure guidance to $130–145 billion by raising the floor from $125 billion, and raised the bottom of its expense range to $165–169 billion. Long-term debt reached $83.664 billion, up from $58.744 billion at the end of December. Every dollar of the revenue sits in two segments, Family of Apps and Reality Labs; there is no compute or AI-infrastructure line, though Mark Zuckerberg told the call the company is 'getting a lot of offers for compute at a significant premium over what we paid for it' and expects to build a large business serving outside customers. Shares fell about 8% in the session and 9.5% by the end of the analyst call.

The Wall Street Journal reported first, on 26 July, that Nvidia is in talks to provide a backstop of roughly $250 billion to help OpenAI lease computing from SB Energy's 10-gigawatt campus at the former Portsmouth Gaseous Diffusion Plant in Piketon, Ohio, and separately to finance $350 billion of OpenAI's chip purchases; Bloomberg reported the same round of talks the following day, taking Nvidia's prospective new deals past $750 billion. None of it is announced or filed. Both accounts rest on people familiar with the matter, describe the negotiations as early and able to collapse, and are confirmed by neither company. Set against Nvidia's own disclosure, the scale is the story: its most recent quarterly filing caps maximum gross exposure across all of its partner facility lease guarantees at $3.5 billion, with $712 million held in escrow behind it, taken in exchange for warrants, carried as credit derivatives and described as immaterial in fair value. A $250 billion guarantee would be roughly seventy times that book, more than a year of Nvidia's revenue and about four times its cash. The mechanism is credit substitution: OpenAI has no investment-grade rating, so a guarantee lets the developer borrow against Nvidia's balance sheet instead of its tenant's. What changes is the instrument rather than the pattern — the September 2025 arrangement recorded here was an equity investment of up to $100 billion in a customer; a guarantee is standing behind that customer's borrowing. Google has moved the same way from the other side, agreeing to backstop lease payments at five data-centre sites for Anthropic in an arrangement Bloomberg says amounts to a $35 billion loan, and SoftBank — which carries more than $130 billion of debt — has committed nearly $65 billion to OpenAI, signing a $40 billion bridge loan to fund it. Ground was broken at the Ohio site on 20 March 2026; a first phase of about 800 megawatts is expected in 2028, and the site's power allocation is controlled by the Commerce Secretary.

Nvidia announced an AI initiative with SK Group worth more than $500 billion, including over two gigawatts of AI data centres on the Korean Peninsula — the first, built by SK Telecom, due to open next year — and said it would invest $1 billion in Naver. The headline figure is gross and runs in both directions: Jensen Huang told Bloomberg Television it counts Nvidia's own purchases of SK Hynix memory as well as SK Group's purchases of Nvidia supercomputers, so it is not $500 billion of Nvidia revenue. Bloomberg counts more than $540 billion of comparable Nvidia deals announced in 2026 alone. What this page records in it is the financing structure: capital committed by the supplier alongside the sale, the pattern first recorded here with the up-to-$100 billion Nvidia–OpenAI investment in September 2025.

Alphabet shares fell more than 7% the day after results — their worst session in over a year, dropping the company back below a $4 trillion valuation — despite a 24% revenue rise and an 82% jump in cloud revenue. A Bloomberg index tracking the Magnificent Seven fell 4.8%, its worst day since the April 2025 tariff announcement, leaving it down 3.7% in 2026 after three years of gains. The Philadelphia Semiconductor Index, up 101% through the first half of the year, lost 17% in July. Year to date Microsoft is down 21%, Meta 9.8% and Amazon roughly flat, while Apple — which has avoided large AI outlays in favour of partnering with model developers — is up 23%. Analyst estimates compiled by Bloomberg now put combined 2026 capital spending by Alphabet, Microsoft, Amazon and Meta at about $724 billion, rising to nearly $950 billion in 2027. For three years, higher AI spending was rewarded by the market; in this week, a company that beat on revenue and cloud growth was sold because it planned to spend more.

In its filed second-quarter results Alphabet reported negative free cash flow of $5.855 billion — the first negative quarter since it went public in 2004 — as capital expenditure doubled year over year to $44.924 billion against $39.069 billion of operating cash flow. The filing's own quarterly reconciliation shows the slide in four steps: $24.461 billion in Q3 2025, $24.551 billion in Q4 2025, $10.116 billion in Q1 2026, then below zero. Alphabet also raised full-year 2026 capex guidance to $195–205 billion from the $180–190 billion given a quarter earlier — the second consecutive raise, against an analyst consensus near $188 billion, and set against $91 billion actually spent in 2025. It was not a weak quarter: revenue rose 24% to $119.8 billion and Google Cloud revenue rose 82% to $24.8 billion, which is GAAP segment revenue rather than an annualized run-rate and therefore the firmest AI-adjacent revenue disclosure on this page. Alphabet still holds $242.5 billion in cash and securities, and trailing-twelve-month free cash flow remains positive at $53.3 billion. The significance is narrower and harder: at the best-positioned hyperscaler, in its strongest cloud quarter, the buildout stopped paying for itself out of the quarter's operating cash.

Alphabet announced $80 billion of equity offerings on June 1 to fund AI compute, then upsized and priced them two days later at $84.75 billion — the largest equity capital transaction ever completed by a listed corporate, according to FactSet. The structure: $15 billion of mandatory convertible preferred depositary shares, $15 billion of Class A and Class C stock, a $40 billion at-the-market programme beginning in Q3 2026, and a $10 billion private placement to Berkshire Hathaway ($5 billion of Class A at $351.81 and $5 billion of Class C at $348.20 a share). The headline overstates the AI portion: Alphabet's own release says ATM proceeds will go 'primarily' to an administrative change in how it meets employee equity-award tax obligations, with about $30 billion of the $40 billion covering 2026 tax obligations — leaving roughly $44.75 billion directed at general corporate purposes including AI capex. The same release disclosed over $85 billion of debt raised in the prior year across six currencies, taking Alphabet's total debt above $100 billion. After the bond market — Meta's $30 billion sale in October 2025 — the equity market.

Anthropic said it had confidentially submitted a draft registration statement on Form S-1 to the SEC for a proposed initial public offering of its common stock. A confidential submission is not a public prospectus: no financials, share count, price range or timetable were disclosed, and the notice was published under Rule 135 of the Securities Act, which is expressly not an offer to sell. The company's last disclosed private mark was $965 billion in May 2026. The significance for this page is procedural rather than promotional — every private-lab revenue figure recorded here has had to be labeled by its disclosure form, because none of it is filed. A completed IPO would put a frontier lab's audited accounts into the public record for the first time.

Leaked audited financials, verified by the Financial Times, showed OpenAI generated $13.07 billion of revenue in 2025 while spending $34 billion — a $20.9 billion operating loss, up from a $3.7 billion revenue / $8.8 billion operating-loss profile in 2024. A headline net loss of $38.5 billion was inflated by a $41.6 billion non-cash charge from its for-profit conversion. Of its costs, $17.2 billion went to Microsoft. It is the cycle's clearest reality check on the gap between reported revenue and the capital being committed — at the epicenter of the buildout.

Anthropic closed a $65 billion Series H led by Altimeter, Dragoneer, Greenoaks and Sequoia at a $965 billion post-money valuation — up from $183 billion the previous September, roughly a 5.3x re-rating in eight months, with $15 billion of the round consisting of previously committed hyperscaler money. The company disclosed run-rate revenue crossing $47 billion. The steepest private-valuation escalation of the cycle, and a central data point for whether these marks are justified.

Amazon said it would invest another $5 billion in Anthropic immediately and up to $20 billion more over time, alongside Anthropic securing up to 5 gigawatts of AWS Trainium chips. As with earlier rounds, the capital is inseparable from a compute commitment — the investment is, in large part, a chip-purchase agreement.

Nvidia reported full fiscal-2026 revenue of $215.9 billion (up 65%), with Data Center revenue of $193.7 billion — its Q4 alone hit $68.1 billion. Data Center went from $115.2 billion the prior year; the company returned $41.1 billion to shareholders. As the AI economy's toll booth, Nvidia's filed, SEC-liable revenue is the single hardest number showing capital actually converting into sales — even as most of it flows from a handful of customers who are themselves financed by Nvidia.

Dueling analyses put AI's macro footprint in sharp dispute. Renaissance Macro's Neil Dutta found AI capex added more to H1 2025 GDP growth than consumer spending; Pantheon's Oliver Allen found that without AI, corporate equipment investment would be negative. But MRB Partners' Prajakta Bhide, adjusting for imported equipment, put AI's contribution at just 40–50 basis points (~20–25% of growth) and rebutted the claim that GDP would have slumped without it; Bespoke found AI was under 5% of GDP. The observable dispute over the same official data is itself the finding.

Elon Musk's xAI closed an upsized $20 billion Series E, above its $15 billion target, with Nvidia and Cisco among strategic investors backing its GPU-cluster buildout. Another mega-round adding to the private capital chasing frontier AI.

SoftBank closed a further $22.5 billion into OpenAI in late December, completing — with $11 billion of co-investors — a $41 billion commitment first struck in March, and giving SoftBank roughly 11% of OpenAI. A rare case of a headline commitment being fully funded rather than merely announced.

Anthropic committed to buy $30 billion of Microsoft Azure compute (plus up to 1 GW more on Nvidia Grace Blackwell and Vera Rubin systems), while Nvidia and Microsoft committed to invest up to $10 billion and up to $5 billion respectively in Anthropic. The same round-trip pattern seen with OpenAI — suppliers funding a customer that buys their products — extended to the industry's second-largest lab, now spanning all three of the biggest labs' compute financing.

On their Q3 earnings calls, Alphabet, Meta, Microsoft and Amazon each raised guidance again, to a combined $380 billion-plus for 2025 — Amazon ~$125 billion, Alphabet $91–93 billion, Microsoft's capex up 45% to $64.55 billion in fiscal 2025. Meta's stock fell 11% the next day as investors questioned the returns. The February guidance of ~$320 billion had been left far behind.

Anthropic agreed to a cloud partnership with Google worth tens of billions of dollars, giving it access to up to one million Google TPUs and over a gigawatt of compute in 2026. The deal is compute capacity, distinct from Google's separate ~$3 billion equity stake — an illustration of how AI 'investments' increasingly mean chips, not cash.

AMD and OpenAI announced a partnership to deploy 6 gigawatts of AMD Instinct GPUs (first 1 GW of MI450 in H2 2026), and AMD issued OpenAI a warrant for up to 160 million AMD shares — roughly 10% of AMD — vesting as deployment scales and AMD's stock hits price targets. A customer being handed equity to become a customer, and a template beyond Nvidia for financing demand with stock.

Nvidia announced it would invest up to $100 billion in OpenAI to fund a 10-gigawatt buildout running on Nvidia systems — the clearest instance of a chip supplier funding a customer that buys its chips. New Street Research estimated every $10 billion invested returns ~$35 billion in GPU purchases. Analysts split sharply: some called it prudent balance-sheet use, others 'bubble-like,' invoking the 1990s telecom vendor-financing bust. The GPUs are to be leased, not sold, sparing OpenAI a depreciation charge and shifting that risk to Nvidia.

OpenAI agreed to a reported ~$300 billion, multi-year purchase of 4.5 gigawatts of additional Stargate data-center capacity from Oracle — one of the largest cloud commitments ever. A purchase obligation of that scale against OpenAI's ~$13 billion in 2025 revenue became a focal point of investor concern about Oracle's own AI bet.

Anthropic closed a $13 billion Series F led by ICONIQ at a $183 billion post-money valuation, more than doubling its prior mark. The company disclosed run-rate revenue rising from about $1 billion at the start of 2025 to over $5 billion by August — the first rung of a valuation climb to $965 billion within eight months.

MIT's NANDA initiative found in 'The GenAI Divide' that roughly 95% of enterprise generative-AI pilots delivered little to no measurable profit-and-loss impact, with only ~5% achieving rapid revenue acceleration — attributing the gap to a 'learning gap' in enterprise integration, not to model quality. Widely cited as evidence the revenue to justify the capex isn't yet materializing; the finding is narrower than the headline, measuring pilots' P&L impact rather than whether AI works.

OpenAI announced a $40 billion round at a $300 billion post-money valuation in partnership with SoftBank, then the largest private funding round on record, to scale compute and research. SoftBank deployed $7.5 billion immediately and the balance later in the year.

Guiding on their Q4 2024 earnings calls, Microsoft, Amazon, Alphabet and Meta signaled combined 2025 capital spending of as much as $320 billion — up from about $230 billion actually spent in 2024. Amazon alone pointed above $100 billion; Microsoft toward $80 billion for AI data centers. It marked the moment AI capex became a macroeconomic force, and the guidance would be revised upward all year.

OpenAI, Oracle, SoftBank and MGX announced Stargate, a new company intending to invest $500 billion over four years in US AI infrastructure for OpenAI, with $100 billion to be deployed immediately. Unveiled at the White House alongside President Trump, it set the scale of ambition for the entire buildout — but the $500 billion is a stated four-year intention, not committed capital.

Amazon added $4 billion to its Anthropic stake, reaching $8 billion total, with AWS as Anthropic's primary training partner and a deep commitment to Amazon's Trainium chips. Like Microsoft–OpenAI, the 'investment' was bundled with a compute commitment — cash and chip purchases moving together.

OpenAI closed a $6.6 billion round at a $157 billion post-money valuation, led by Thrive Capital with Microsoft, Nvidia, SoftBank and others participating — up from a reported ~$80 billion earlier in 2024 and ~$29 billion in 2023. The first rung of a valuation climb that would reach $500 billion within a year.

Weeks after ChatGPT's launch, Microsoft announced the third phase of its OpenAI partnership — a multiyear, multibillion-dollar investment officially confirmed only as such, and widely reported at about $10 billion (Bloomberg, Semafor) though never officially confirmed. Following prior investments in 2019 and 2021, it lit the fuse on the hyperscaler capital-spending race that defines this page.

Microsoft invested $1 billion in OpenAI under an exclusive computing partnership, with Azure becoming OpenAI's sole cloud provider. The structure — capital paired with a commitment to buy the investor's compute — became the template later echoed across the industry's biggest AI deals.