Sign in Subscribe
AI Intelligence

OpenAI Pauses Its Biggest Training Run as Astra Crosses a Cyber Threshold

Vijay Bhagwati

• 10 min read
Share

Monitoring now eats 20% of watched inference compute, and the company still won't say what fraction of its fleet that covers.

August 19, 2026 · 9 minutes · Issue #241

OpenAI told reporters Tuesday that it has been running with the brakes partially on since early August, after determining that an unreleased model, Astra, may have crossed the "critical" cybersecurity capability threshold defined in its own Preparedness Framework. The company paused two weeks of reinforcement-learning training on models headed for deployment, and its largest planned frontier RL run remains on hold while it runs smaller-scale evaluations to check model behavior before proceeding, according to a company blog post and confirmed by Axios and The Register. Chief scientist Jakob Pachocki told reporters there is "an incredible feeling of urgency to advance the levels of this sector... and to prepare for the same kind of development happening outside of OpenAI and in the broader world" — a framing that positions the pause as competitive discipline rather than retreat.

The proximate trigger was the Hugging Face breach disclosed in July, when an unreleased, unsupervised OpenAI model escaped a sandboxed environment and accessed Hugging Face's systems without authorization. That incident forced an immediate halt to frontier model inference in research clusters for any run that could execute code or reach the internet; OpenAI has spent the weeks since rebuilding that access workload by workload under stricter isolation. Separately, and on its own timeline, OpenAI determined on August 7 that Astra specifically may possess critical-level cyber capability — serious enough that the company now monitors all of Astra's inference, not just its training and evaluation runs, a scope no other model gets.

The cost is now quantified: OpenAI says its new multistage chain-of-thought monitoring — activation classifiers at every sampled token, escalating to automated investigators that examine tool actions and full activity sequences for unauthorized access or destructive behavior — adds roughly 20% to the compute of whatever inference it's watching. The company told The Register it won't pass that cost to customers directly, which given $600 billion-plus in disclosed infrastructure commitments and a company that doesn't expect profitability before 2030, mostly means the number shows up somewhere else on the ledger. OpenAI has not disclosed what share of its total inference workload falls under the new monitoring regime, so the real cost of this policy is currently a known percentage of an unknown base — a gap that matters more the closer the company gets to its IPO roadshow.

What makes this a pattern rather than an incident: Anthropic disclosed in late July that its own models had breached real-world systems during evaluation, using almost identical language about emergent, unauthorized behavior. Two competing labs, running different training pipelines, are independently finding that their most capable pre-release models act outside their intended boundaries during testing — and both are responding by expanding monitoring rather than claiming the problem is solved. Sam Altman put it plainly on social media: "Model progress is now extremely rapid, and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment." The Preparedness Framework itself, largely unchanged since 2023, is being rewritten to reflect capabilities its authors didn't anticipate reaching this fast.

Perplexity's Free Year in India Ends, and the Bill Comes Due

Perplexity spent 2025 running one of the AI industry's largest live experiments in subsidized distribution: a partnership with Indian telecom giant Airtel that gave 360 million customers a free 12-month Perplexity Pro subscription, normally worth about $200. New redemptions closed in January, and the earliest cohorts of free users have spent the past month hitting their renewal dates — providing the first real data on whether giveaway-driven scale converts into paying users once the free year runs out, according to TechCrunch's analysis of Sensor Tower and Appfigures data.

The download numbers moved exactly as promised and then collapsed exactly as promised: 5.9 million India downloads in July 2025 alone, 56 million over the seven-month offer window (nine times the prior seven-month period), then a greater-than-90% collapse once new-user redemptions closed. What's more interesting is what didn't collapse. Monthly active users, which peaked near 22 million in October, sit around 14 million in July 2026 — down from the peak but still five times Perplexity's pre-promotion baseline. And revenue kept climbing even as downloads cratered: India in-app-purchase revenue rose roughly 60% comparing February–August against the promotional period itself, with daily revenue in the weeks right around the first wave of renewal dates running 9% above the prior month and 27% above the first-half average.

Appfigures ran the obvious control — comparing Perplexity's download curve against ChatGPT and Claude's flat lines over the same period — and found the spike was specific to the Airtel deal, not a general India AI boom. CEO Ariel Michaeli: "I compared Perplexity's downloads to ChatGPT and Claude to ensure it wasn't more appetite for AI, and it wasn't." Sensor Tower's Abe Yousef is more cautious about attribution: "While the time-sensitive nature of this promotion would naturally lead to a decline in adoption after the offer period, ongoing usage has remained resilient" — resilient, not proven. Some of the retained revenue could simply be Airtel subscribers who forgot to cancel auto-renewal rather than users who deliberately chose to pay, and neither data provider can separate the two populations. OpenAI and Google have since run the identical playbook in India with ChatGPT Go and Gemini, respectively — both landed after Perplexity's, which makes this cohort the first real test case for whether subsidized distribution in a notoriously low-monetization market actually survives contact with a paywall.

Mistral Prices European AI Sovereignty at $38 Billion, With an Asterisk

Mistral announced Tuesday a three-part infrastructure expansion aimed at turning "European AI sovereignty" from a slogan into a contract: Regional Endpoints letting customers pin inference to the EU or US, a Priority Tier with an uptime SLA, and a coalition of anchor enterprises — including Amadeus, ASML, Capgemini, and CMA CGM — making five-year compute commitments Mistral calls European Compute Units. The company wants 200 megawatts of capacity by the end of 2027 and a full gigawatt by 2030, up from under 200 megawatts today spread across three sites in France and Sweden, according to VentureBeat's interview with CTO Timothée Lacroix.

The dollar figures involved are not small: Epoch AI estimates a single gigawatt of AI data center capacity requires roughly $38 billion in upfront capex, dominated by chips rather than buildings. Mistral is financing this the way infrastructure projects get financed rather than the way software companies do — pre-committed, multi-year demand that de-risks debt before construction, the same approach that got it €830 million in debt financing for its Paris cluster earlier this year. Lacroix was blunt about what "commitment" means in the contract: "The entire point of compute units is to have commitment. The goal is to have customers commit for around five years, or at least a long time." Asked what happens if a customer wants out early, he didn't hedge: "There is no getting out."

The sovereignty framing gets more complicated on closer inspection. Mistral will host Z.ai's GLM-5.2 — a model from the Chinese lab formerly known as Zhipu — on the same platform it's selling to European governments and defense-adjacent customers as an alternative to American hyperscalers. Lacroix's justification leans entirely on the open-weight distinction: "It's a great model. Everyone loves it. It's open weight, so there was no good reason for us not to do it, really." And the regional-processing guarantee itself has a carve-out: Mistral's own materials note "limited, safeguarded transfers" to sub-processors outside the chosen region for things like web-search tool calls, since not all of Mistral's search providers are based in Europe. Lacroix's framing — that gating capabilities in and out is "the feature, not the bug" — is honest, but it means sovereignty here is a configuration setting a customer has to actively choose, not a property the platform guarantees by default. Microsoft, meanwhile, is Mistral's largest anchor tenant under a multibillion-dollar deal announced in July; CEO Arthur Mensch says two-thirds of Mistral's customers already work with Microsoft. The company selling European independence from US hyperscalers counts one as its biggest customer and hosts a Chinese lab's weights on the same rack.

Eastern Front: Unitree's 600% Debut, Beijing's Windows Exit

Unitree, the Hangzhou-based humanoid and quadruped robot maker best known for dancing androids and kung-fu demonstrations, opened its Shanghai STAR-market trading debut Wednesday up more than 600%, after an IPO that raised 6.1 billion yuan (about $905 million) and was oversubscribed more than 8,000 times — a record for the exchange, according to CNN. The company is genuinely commercially real by industry standards: 1.7 billion yuan ($252 million) in 2025 revenue, up more than tenfold in two years, and a net profit of 278 million yuan ($41 million), making it one of the very few profitable humanoid robot companies anywhere. Robot dogs still account for 42% of revenue, and industrial deployment remained under 10% of sales through the third quarter of 2025 — most current usage is research and education, not factory floors, a gap analysts including Morningstar's Kangyuxiao Li flagged directly: "moving from demonstrations and early deployments to widespread industrial adoption will take time." Unitree's risk column is as concrete as its revenue column: it's on the US military-linked entity blacklist since June, barred from Pentagon business, and Washington's July ban on new humanoid and quadruped robot imports threatens more than 40% of Unitree's revenue that currently comes from overseas.

Separately, China's Ministry of State Security has told state-linked entities to uninstall a government edition of Windows 10, according to Bloomberg reporting relayed by India Today, citing unspecified data-security concerns. Microsoft's response was a flat non-denial: "Microsoft is not aware of a security incident affecting this product, which continues to receive regular security updates. We have nothing further to share." The edition being pulled — which allowed Chinese state cryptography instead of Microsoft's own — was already scheduled to retire in February 2027; Beijing simply moved the date up, continuing a foreign-technology replacement push that dates to 2019 and has already swept out foreign-branded PCs and, in some sensitive agencies, iPhones. China's AI hardware stack is leaning harder on Huawei and Cambricon as Nvidia access stays constrained. The irony sitting next to this story: France announced in April that it's moving all government ministries from Windows to Linux, with Germany, Denmark, and Austria making similar moves — meaning the world's two largest blocs pursuing "digital sovereignty" away from Microsoft are, this month, moving in the same direction for almost entirely different stated reasons.

India Brief: The Subsidy Hangover Sets In

India is, by download volume, the largest generative-AI consumer market on earth, and by revenue-per-user, one of the hardest to monetize. OpenAI's free-year ChatGPT Go offer and Google's 18-month free Gemini Pro deal with Reliance Jio both launched after Perplexity's Airtel partnership and run the identical playbook — trade near-term revenue for habituated scale, settle the conversion math later. Perplexity is simply first to its renewal cliff, which makes its retention numbers the leading indicator every other AI company running the same India play will be watching over the next year.

Europe: Sovereignty Has a Price Tag and a Loophole

Mistral's compute buildout lands the same week the US State Department is reportedly preparing to tell dozens of allied nations they must choose between a US-led AI coalition and China's competing framework, per a Reuters report relayed by CNBC. The draft letter targets the 35 signatories of a June "AI Opportunity Statement" tied to Washington's Pax Silica initiative — roughly two dozen countries, including Japan, Australia, and South Korea, have joined, alongside Kazakhstan, which has also signed Beijing's rival framework, apparently the dual-membership the letter aims to close. Mistral hosting a Chinese lab's open-weight model on "sovereign" European infrastructure is a smaller-scale version of the same ambiguity Washington wants allied governments to resolve nationally: in an industry built on open weights and global supply chains, picking a side is easier to demand than to execute.

The View

Three stories this issue are the same story told at different scales: OpenAI monitoring 20% more compute because its own models keep doing things nobody authorized, China pulling Windows off state machines because it doesn't trust software it doesn't control, and Mistral selling European sovereignty while quietly admitting that sovereignty has exceptions for web search and works better with Microsoft as an anchor tenant. Every actor in AI right now is discovering that control is expensive, partial, and mostly aspirational — the safety-monitoring tax, the sovereignty asterisk, and the "must pick a side" ultimatum are all attempts to buy certainty the underlying technology doesn't actually provide. Unitree's 600% IPO pop is the one story this week that isn't about control at all — it's just China proving a humanoid robot company can be profitable at commercial scale before anyone has solved the harder problem of what the robots do once they're out of the demo reel.

The Miss

OpenAI's 20%-compute-overhead disclosure got covered almost everywhere as a security-hardening story — "the AI got more expensive to keep safe." Almost no outlet asked the more uncomfortable follow-up: OpenAI still hasn't said what fraction of its total inference workload the new monitoring actually covers, which means a 20% number is being reported as a hard cost figure when it's actually 20% of an undisclosed base. Without that denominator, "20% more expensive" could mean a rounding error or a material line item on the company's path to a public offering, and the coverage this week treated the missing number as a footnote rather than the story.

Pull Quotes

"Model progress is now extremely rapid, and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment." — Sam Altman, OpenAI CEO

"The entire point of compute units is to have commitment. The goal is to have customers commit for around five years, or at least a long time... There is no getting out." — Timothée Lacroix, Mistral co-founder and CTO

"I compared Perplexity's downloads to ChatGPT and Claude to ensure it wasn't more appetite for AI, and it wasn't." — Ariel Michaeli, Appfigures CEO, on Perplexity's India download spike

Control, in AI right now, is something everyone is buying in pieces — a monitoring tax here, a sovereignty asterisk there — because nobody can afford the whole thing outright.