What happened in AI yesterday — Monday, August 17, 2026.

The Read

Monday delivered the largest private compute deal ever assembled, and the interesting part is who is standing behind whom. At PORTS-Pike — a decommissioned uranium enrichment plant on federal land in southern Ohio — SB Energy will build 8 gigawatts, OpenAI will be the sole tenant on a 20-year lease, NVIDIA will be the exclusive chip supplier, and NVIDIA will also guarantee roughly $105 billion of the rent and power payments if OpenAI cannot pay. The chip vendor is underwriting the tenant's ability to buy its chips. Separately Anthropic disclosed $11.5 billion of second-quarter revenue, up fourteenfold year over year and adjusted-operating profitable, putting its run rate past $65 billion against OpenAI's roughly $40 billion. And Greg Brockman wrote the sentence OpenAI has been circling for a month: 'we underestimated the real-world cyber capabilities of our AI models.' The day-zero read: the revenue is real, the buildout is real, and the financing is increasingly circular. Hold all three thoughts at once.

🌊 Tide — the megatrend layer

No shift. All four tides hold. One confirmation on governance-as-market-structure, and it is a new role rather than a new rule: the US Department of Energy supplied the land for the largest private compute project ever announced — the former Portsmouth Gaseous Diffusion Plant in Pike County, Ohio — with OpenAI as sole tenant, SB Energy as owner-operator, and $84M in Codex credits attached for roughly 844,000 Ohio college students. This tide has logged the state as rule-writer (export controls, the EU AI Act), as funder (EuroHPC gigafactories), as evaluator (UK AISI, TRAINS), and as producer (DOE's Genesis open models). Landlord is new. The federal government is now a counterparty in the capital stack of frontier compute rather than a regulator standing outside it, and the community-benefit package attached to the site is what that leverage buys.

DOE land hosts an 8GW OpenAI campus at the former Portsmouth enrichment plant, with student compute credits attached

OpenAI: OpenAI joins the PORTS-Pike project

🌊 Waves — weeks to quarters

PORTS-Pike: NVIDIA guarantees the rent so OpenAI can buy NVIDIA chips

OpenAI announced an agreement with SB Energy, NVIDIA and the Department of Energy for roughly 8 gigawatts-IT in Pike County, Ohio. SB Energy builds, owns and operates under a 20-year lease with OpenAI as sole tenant; the site hosts NVIDIA compute exclusively; first 800MW online in 2028, six-year buildout to 2032, 35,000 construction jobs and 2,500 permanent. Jensen Huang confirmed NVIDIA is backstopping about 4.25GW — phasing in as buildings enter service between 2028 and 2030, covering 'defined portions of lease and power payments, along with a specified residual-value commitment — not the full cost of the site' — with an option on the remaining 3.75GW. Reported guarantee figures cluster around $100–105B, down from the $250B the WSJ floated, and The Information reports NVIDIA is trying to cap its exposure at 25% of total financing. NVIDIA is also investing $1.5B in SB Energy, half at signing and half into SB Energy's IPO next month, down from a reported $3B. SB Energy and SoftBank will build 10GW of new generation, 9.2GW of it gas, plus $4.2B of grid work with AEP Ohio. Huang's own framing of the return: each generation of AI factory systems at the site could mean roughly 1.5 million GPUs and $150–200B of NVIDIA revenue.

Roadmap implication: this is vendor financing at national-infrastructure scale, and the structure deserves the scrutiny the headline number will not get. NVIDIA books revenue on chips whose purchase it partially guarantees, at a site whose rent it partially guarantees, for a tenant whose ability to pay depends on demand that has not yet materialised in 2028. That can be entirely sound and still be circular. If you model vendor health, note that Huang's careful wording — defined portions, residual value, not the full cost — is doing a lot of work, and that the guarantee phases in years from now on terms nobody outside has seen.

OpenAI: OpenAI joins the PORTS-Pike project
Data Center Dynamics: NVIDIA confirms it will back 4.25GW at PORTS-Pike
CNBC: NVIDIA financing OpenAI's Ohio data center

Anthropic's revenue rose 14x to $11.5B in Q2 — and it passed OpenAI on run rate

Anthropic's second-quarter revenue reached $11.5 billion, up roughly fourteenfold from $787 million a year earlier and more than double the $4.73 billion it posted in Q1 — more than its entire 2025 revenue, in one quarter. It was profitable on an adjusted operating income basis. Run-rate revenue has passed $65 billion, with some estimates above $70 billion, against roughly $40 billion annualized at OpenAI; in July 2025 Anthropic's run rate was $4 billion. The disclosure comes as the company briefs investors ahead of a public offering that could come as soon as next month, at a valuation its backers have told the FT they expect above $2 trillion. As recently as Q1, OpenAI was the larger of the two on revenue.

Roadmap implication: the enterprise-first, coding-heavy strategy has now overtaken the consumer-first one on run rate, which reframes every vendor-concentration conversation you are having. Two consequences worth acting on. First, Anthropic's negotiating position hardens into the IPO, so lock terms before the listing rather than after. Second, watch The Information's reporting that both labs are building industry-specific applications — a vendor growing this fast on enterprise revenue eventually meets its own customers in the market.

The Information: Anthropic revenue jumped 14 times in the second quarter

Brockman: 'we underestimated the real-world cyber capabilities of our AI models'

Greg Brockman published 'The Defender's Window,' OpenAI's post-mortem manifesto on the Hugging Face incident. The concession is unusually plain: an agentic collective 'was able to autonomously penetrate not just OpenAI research infrastructure but also the production infrastructure of another company,' and 'we are strengthening our safety requirements accordingly.' He illustrates the defensive case with a small demonstration — asking ChatGPT Work to audit his own site turned up 13 issues in about fifteen minutes, fixed within the hour. The forward-looking passage is the one to note: Brockman flags that Z.ai's GLM-5.3 open weights 'appear slated to be released at the end of August' and 'seem likely to significantly accelerate the threat landscape.' That also corroborates that Z.ai's two-week safety-hardening delay is still holding.

Roadmap implication: a frontier lab is now publicly forecasting a competitor's open-weight release as a dated threat event. Whatever you think of the messenger's incentives, the date is actionable — if GLM-5.3 weights land at end of August with 84.5% CyberGym performance and no gating, the capability floor for anyone attacking you moves that week. Get your patch cycle and external attack surface review scheduled ahead of it, not after.

OpenAI: The Defender's Window

AI borrowing is now moving the cost of capital for everyone

Hyperscalers, data-center developers and AI infrastructure projects have issued roughly $308 billion of debt through July, nearly 13x the same period last year. Because data centers are long-lived assets, much of that paper is long-dated and competes directly with 10- and 30-year Treasuries. The real market yield on 30-year Treasuries now stands at 3.00%, the highest since the Great Recession. Monitor the Situation frames it as crowding-out in reverse: not government borrowing squeezing private credit, but private AI borrowing lifting the government's cost of capital — alongside the deficit, inflation and the Fed. It pairs this with the Chow–Halperin–Mazlish argument that expectations of transformative AI should raise real rates regardless, since anyone expecting to be much richer later saves less now.

Roadmap implication: AI capex has become a macro variable, which means it is now in your cost of debt whether or not you buy a single GPU. If you are financing anything long-dated in the next eighteen months — a facility, an acquisition, a refinancing — the AI buildout is part of your rate. Two counterweights worth holding: this is one plausible reading among several for why rates are high, and the Wood Mackenzie finding that ~72% of interconnection requests are phantom suggests some of that $308B is funding capacity that never gets built.

Monitor the Situation, 8/17: PORTS-Pike and treasury yields
Chow, Halperin & Mazlish: Transformative AI, existential risk, and real interest rates

🌊 Ripples — actionable within days

Stripe finalizes OpenRouter — and the revenue multiple depends on which number you believe

The Information and Axios both confirmed Monday that Stripe has finished its deal for OpenRouter at more than $7 billion, against a $1.3 billion valuation months earlier. Neither company has confirmed on the record; Axios says an announcement is expected soon. The multiple is genuinely unclear from public reporting: The Information's Monday briefing cites $50 million annualized as of April, up fivefold from October, while its own July reporting put the figure near $140 million annualized. That is the difference between roughly 140x and roughly 50x revenue. OpenRouter has raised $164 million total from CapitalG, a16z, Menlo, NVentures, ServiceNow Ventures, MongoDB, Snowflake and Databricks.

Do this now: when someone quotes you an AI comparable, ask which revenue number and which month. A 3x spread in the denominator on the most-watched deal of the month is a useful reminder that these multiples are being repeated far more confidently than they are known. If OpenRouter sits in your stack, the integration window is when to get data portability and routing-policy terms in writing.

Axios: Stripe, OpenRouter and PayPal
Bloomberg: Stripe nears deal to buy OpenRouter for over $7 billion

Cursor ships Origin — a GitHub competitor, two days after becoming part of SpaceX

Cursor released Origin in early beta for paid users: code hosting with repos, pull requests and code browsing, and sync with existing GitHub repositories, with more agent-native features promised. The timing is the story — SpaceX completed the $60 billion all-stock acquisition on Friday, Cursor staff are being folded into SpaceX's AI division rather than kept as an independent unit, and the first product out of the gate attacks GitHub's core business. GitHub shipped Copilot in October 2021, eighteen months before Cursor existed.

Do this now: if you are a GitHub Enterprise shop, put Origin on the evaluation list for one team — not to migrate, but to learn what agent-native code hosting looks like before your developers ask. The strategic read is that the coding-agent vendors are moving up into the repository layer, where the switching costs and the training data both live.

Monitor the Situation, 8/17
The Information: SpaceX completes $60 billion Cursor acquisition

Google buys a dead airline's internal data for $10M to train on

Google paid $10 million for Spirit Airlines' corporate data, outbidding a $7.5 million offer from Mercor. Spirit ceased operations on May 2, 2026. The purchase covers emails, Teams messages, spreadsheets and calendars, and is intended for AI training. It is one of the cleanest examples yet of a new asset class: the internal communications exhaust of a defunct company, sold for model training because it is the kind of messy real-world enterprise workflow data that public web scraping does not contain.

Do this now: ask your general counsel what happens to your internal corporate data in a wind-down, an acquisition, or a bankruptcy — because it is now a bid asset with a market price, and most contracts and employee-communication policies were written when it was a liability instead. If you have ever offloaded a subsidiary, the question is not theoretical.

Monitor the Situation, 8/17

Higgsfield raises $400M at $5.4B on $700M of annualized revenue

DST Global led Higgsfield's Series B, with Goldman Sachs Alternatives, Tribe, Smash, Fifth Wall, Valor, Intel Capital, Liberty Global, Mirae and NTT DOCOMO Ventures joining existing backers Accel and Menlo. The valuation quadrupled in eight months from $1.3 billion. The company claims $700 million annualized revenue — up from roughly $20 million a year ago — with 30 million-plus users across 238 countries and 390 of the Fortune 500 as customers. Founder Alex Mashrabov named the constraint directly: 'Video is one of the most compute-intensive domains in AI.'

Do this now: a 35x revenue increase in twelve months in AI video is a signal about demand, not about Higgsfield. If your marketing organisation still treats video as the expensive channel it budgets around, that assumption has a shelf life measured in quarters. Run one campaign through a generative video tool this quarter purely to recalibrate what it costs.

TechCrunch: Higgsfield raises $400M Series B at $5.4B
PR Newswire: Higgsfield Series B

OpenAI funds 14 AI-policy projects — and mentions in passing that ChatGPT has a billion users

OpenAI awarded $1 million in cash plus up to $1 million in API credits across 14 independent policy projects, from more than 400 applicants, spanning the US, EU, Brazil, Singapore and South Korea. Grantees include AEI and Urban on AI and the American workforce, ECIPE on a 'Right to AI' framework, the Abundance Institute on state-by-state energy and a Data Center Atlas, the Progressive Policy Institute on 'livelihood insurance,' the Institute for Security and Technology on measuring uncontrolled recursive self-improvement, and the Nuclear Threat Initiative on AIxBio cross-border information sharing. Six-month projects, results in 2027. The post also states, without ceremony, that ChatGPT is 'used by 1 billion people' — six days after Google announced Gemini crossing the same mark.

Do this now: read the grant list as a map of where OpenAI expects the policy fight to land — energy siting, labour displacement, and recursive self-improvement measurement. Those are the three files to have a position on before someone asks you for one. And note the billion-user line: two assistants at a billion users each, disclosed six days apart, is the distribution story arriving in the footnotes.

OpenAI: New policy ideas for the Intelligence Age

Read this and every past edition at excelsiorgroup.ai/insights/signal.

The Signal — The Excelsior Group. Tide, wave, ripple: sorting what actually matters from what merely happened.