Aleth

Open access research

ALETH / AI-BRIEF / 2026-08-22 / OPENAI PAUSED TRAINING ON ITS NEWEST MODELS

OpenAI paused training on its newest models

Also this week: Anthropic passed OpenAI on revenue, Stripe agreed to buy OpenRouter & Unitree's IPO valued the humanoid robot maker at $50bn.

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

Browse by day:

Monday 17 August

Nvidia guaranteed up to $105bn of OpenAI’s Ohio data centre lease.

OpenAI has contracted for about 8 gigawatts of IT load at SB Energy’s PORTS-Pike campus in Pike County, Ohio, on a 20-year lease, at the former Portsmouth gaseous diffusion plant. Nvidia’s 8-K caps its payment obligation at $105bn for the initial 4.3 gigawatts of IT load, payable if OpenAI goes insolvent or stops paying, and Nvidia is putting a further $1.5bn into SB Energy. The site will exclusively host Nvidia compute, and the first 800 MW is expected in 2028.

Jensen Huang rejected the circular-financing reading: “Is this circular financing? No. OpenAI will pay the lease.“ The guarantee runs for 20 years from each lease commencement and falls away if OpenAI reaches a satisfactory credit rating.

Amazon is buying rare books in bulk, cutting off their spines and destroying them after scanning them for AI training data.

404 Media put a tracking device in a shipment of rare books and followed it to a warehouse in Las Vegas, where the team that works on the scanning is known internally as VGT3. Amazon said it “purchases books through commercial channels to improve the products and services customers use“.

An autonomous security agent built by Wiz found and exploited a flaw in a Snowflake repository that GitHub’s own scanner had cleared.

The injection went live when a pull request was merged on 18 June and was discovered five days later, and the agent used it to exfiltrate a credential that gave read access across Snowflake’s engineering, security compliance and bug bounty tracking projects in Jira. Copilot reviewed the merged change and reported it clear. Snowflake patched it the same day it was reported.

Google won a bankruptcy auction for the data of collapsed airline Spirit with a $10m bid, to train its models on it.

The lot runs to 100 million emails, 500 million Microsoft Teams records and about 30 million lines of code, and excludes 97.5 million passenger profiles. Mercor was the underbidder at $7.5m. A flight attendants’ union then objected that decades of disciplinary, payroll and training records sit inside the tranche, and the sale hearing was put back to 9 September.

The US Department of Justice is reportedly investigating whether Andreessen Horowitz partners sit improperly on the boards of competing AI companies.

The seats under scrutiny are Ben Horowitz at Databricks and Martin Casado at Fivetran, and the statute is Section 8 of the 1914 Clayton Act. Venture firms have escaped interlocking-directorate cases partly because startups fell below the Act’s size thresholds, which AI companies’ valuations now exceed.

Tuesday 18 August

OpenAI paused training on its most capable unreleased models.

The lab stopped reinforcement learning on its newest deployment-bound models for two weeks while it hardened its research environments, and its largest planned frontier run is still on hold. Sam Altman told Time the decision followed several research observations showing varying degrees of misalignment: “I think it is a good time to slow down“. He added: “Getting AI safety right is more important than any company’s momentum“. No date has been given for Astra’s release.

The Financial Times reported that OpenAI closed its Preparedness team, which judged whether its models could cause catastrophic harm, at the end of July and handed its work to existing groups. OpenAI denies disbanding it.

Claude designed protein binders against 14 of 15 targets, with hit rates of 22% to 35% where 10% to 15% is normal.

Anthropic ran Mythos Preview and Opus 4.8 autonomously in Claude Science, with up to 12,500 H100 hours in multi-target mode, and had the designs made and tested by two external labs. Some binders matched or beat the best published affinity for their target. A generally available model, Opus 5, separately processed raw NMR and LC-MS files in 23 and 19 minutes and matched a contract lab’s purity reading at 96.4% against 96.3%. Protein design stays blocked in Anthropic’s most capable model, and the company says an access programme for scientists is one of its highest priorities.

Anthropic’s revenues passed OpenAI’s in Q2, earning $11.6bn vs $6.7bn.

OpenAI grew 18% quarter on quarter (QoQ) to $6.7bn and its operating margin fell further into the red. Anthropic roughly doubled to $11.6bn and posted a small operating profit. It is the first reported time Anthropic has out-earned OpenAI. On the Vercel gateway, Anthropic also earns far more per use than its rivals. Anthropic’s annualised run rate passed $65bn at the end of July, seven times its level a year earlier, according to Bloomberg sources.

Etched raised $700m at a $21bn valuation led by Jane Street, and shipped its first rack to the same firm.

The valuation doubled in 26 days, from $10.3bn on 23 July. Jane Street tested the hardware before leading, and four proprietary trading firms now back Etched.

Z.ai put GLM-5.3 on its API at GLM-5.2’s prices, $1.4/m input and $4.4/m output.

Artificial Analysis scores it 60 on its Intelligence Index, level with Kimi K3 atop the open-weights field and seven points above GLM-5.2, though it puts the cost per index task at $0.68 against $0.44 because the model is wordier. The weights are not yet out.

China let ByteDance and Tencent each import about 10,000 Nvidia H200s.

US rules already allow up to 100,000 H200s per company, and buyers still need sign-off from China’s state planner, the NDRC. The H200 is at least two generations behind Nvidia’s best, and the Financial Times reported more Chinese firms could follow.

Situational Awareness offered part of its Anthropic stake at a 20% discount on a 12-hour deadline, then told the buyers it had a better deal with Citadel.

Anthropic holds approval rights over transfers of its shares, so Leopold Aschenbrenner’s fund could only approach firms that already held them. While that ran, it was negotiating with Citadel and Millennium, and Citadel won the leveraged book in the early hours of 30 July at about 10% below market. Prime brokerage reports showing tech leverage falling told the market someone large was in trouble before any of it was announced.

Wednesday 19 August

Stripe agreed to buy OpenRouter.

OpenRouter routes requests across >400 models from >80 providers, processes >10 trillion tokens a day and counts >10 million developers and companies. Neither company disclosed terms. Patrick Collison said: “Tokens are the central currency for companies building with AI.“ OpenRouter’s Alex Atallah said routing decisions will stay driven by what is best for the user, and the deal is subject to customary closing conditions.

Five US agencies warned of AI-written exploits against Siemens controllers.

The NSA, CISA, FBI, Department of Energy and Environmental Protection Agency said attackers are using AI-generated scripts, disguised as monitoring tools, against internet-exposed Siemens S7 series programmable logic controllers (PLCs) in critical manufacturing, energy, water and wastewater, chemical, food and agriculture and commercial facilities. The agencies describe persistent reconnaissance and capability development, and say AI assistance dramatically reduces the expertise and time needed to produce working industrial control exploits.

Marvell gave Google a warrant over 59.0m shares in a custom silicon deal.

The exercise price is $206.6, valuing the warrant at about $12.2bn if fully exercised. Vesting is almost entirely tied to purchases: 240 tranches, one for each $500m of custom product revenue through fiscal 2033, against a commercial agreement signed on 29 July covering inference accelerators, storage and network controllers and near-memory compute attached to Google’s TPU ecosystem. Marvell shares rose nearly 8% and Broadcom, until now Google’s main custom chip partner, fell >5%.

UK chip startup Fractile is in talks to raise about $600m at a $6.5bn pre-money valuation after an Anthropic supply deal.

The initial agreement covers roughly $250m of inference chips that will not be ready for use until 2027. The London company was valued at around $1bn in May, when Accel, Founders Fund and Factorial Funds led a $220m round.

OpenAI’s CFO told staff the company will be public in “2027, or sooner”.

Sarah Friar said at an all-hands that the IPO is “not a finish line, it is a milestone, another fundraise“. She showed slides putting OpenAI’s revenue run rate up 35% quarter to date, its enterprise revenue run rate up 50% and its coding and work products at 20 million weekly active users. Ramp’s data puts OpenAI ahead of Anthropic on QoQ growth in business API spending in Q3, 82% against 76%.

Unitree’s $905m IPO sends the humanoid robot maker to a $50bn valuation.

Shares in China’s leading humanoid robot maker opened 629% above their IPO price and closed up 460%, valuing the company at around $50bn. The STAR Market offering was oversubscribed more than 8,000 times, a record, with DeepSeek investing $20.8m alongside backers including Tencent and Alibaba. Unitree shipped more than 5,500 humanoid robots in 2025 and generated $252m in revenue, making it one of the few profitable companies in the sector. However, few of its robots are yet deployed commercially, with sales still concentrated in research and education. At nearly 200 times annual sales, the valuation reflects enormous expectations for embodied AI.

OpenAI patched Codex after GPT-5.6 Sol deleted real user files while running.

A cleanup command meant for temporary folders could point at a real home directory when the model used system variables such as $HOME. Codex now verifies deletion targets first, creates fresh temporary folders and blocks accidental full-access mode.

Thursday 20 August

London’s Callosum raised a $100m seed round, the first investment made by the UK Sovereign AI Fund.

Atomico led, with Plural and DCVC alongside the state fund, which the government launched in April with £500m. Callosum’s software matches tasks to the cheapest suitable model and chip. AI minister Kanishka Narayan said success will depend on using chips as efficiently as possible.

Nvidia pays Poolside $6bn for a non-exclusive licence to its model-building software, hiring 109 staff and investing $1bn at a $12bn pre-money valuation.

The investor letter says the deal is “not an acquisition and it is not an acquihire“, and the three founders stay. It is Nvidia’s third licence-and-hire deal. The letter also sets out why Poolside stopped building frontier models: it had six weeks to raise $2bn for a 40,000 GB300 cluster due in January, missed it and lost the cluster.

Anthropic will let enterprise customers keep the 30 days of data its safety system needs in their own cloud.

The 30-day requirement came in with its Mythos and Fable models in June. Anthropic found that regulated customers could not use models whose provider held their logs, and built the alternative with >100 of them. A day earlier OpenAI previewed Private Safety Processing, which flags misuse across sessions while data stays in infrastructure the customer controls or under its own keys.

A US judge acquitted ex-Google engineer Linwei Ding of economic espionage.

A San Francisco jury had convicted him in January on seven counts of trade secret theft and seven of economic espionage over 1,255 documents uploaded from spring 2022 to spring 2023. Judge Vince Chhabria found too little evidence that Ding intended to benefit the Chinese government at the time of the uploads, and left the trade secret convictions standing. Sentencing is set for 1 September.

Friday 21 August

Anthropic expects its IPO to match or beat SpaceX’s record $86.2bn listing, and could file publicly by the end of August.

The New York Times reported that its bankers are floating a raise of >$100bn. Anthropic’s revolving credit facility is also expected to exceed its roughly $10bn target as banks bid for roles on the listing.

OpenAI asked California to tighten the AI law it once opposed.

OpenAI said SB 53 should be amended to require monitoring of frontier models while they are still in training or evaluation for incidents that could bypass a third party’s security controls, and to strengthen cybersecurity throughout development. OpenAI calls the approach reverse federalism. California’s legislative session is in its final days.

Born on Substack · read and comment there