August 19th 2026
Curated AI news and stories.
Unitree closed its Shanghai debut up 460%, then gave a fifth of it back
Unitree opened its first day on Shanghai's at 1,100 yuan, up 629% on the 150.80-yuan offer, and spent the rest of the session handing it back, closing at 845 yuan, up 460%. The close values China's first mainland maker near 342 billion yuan, about $48 billion, roughly 5.6 times its IPO and, by ts2.tech's math, near 857 times projected 2026 earnings. The people holding the risk are retail: the tranche was 8,288.82 times oversubscribed, a single winning lot showed about 347,000 yuan of paper profit at the close, and founder Wang Xingxing keeps 68.8% of the vote. The fade from open to close is the number that matters. A stock that opens up 629% and closes up 460% is priced by the crowd that queued for the lottery, not by anyone valuing a company that shipped fewer humanoids in the first half than AgiBot did. Meituan's 420-million-yuan stake is now worth about 38.6 billion. Caixin counts 30 to 50 Chinese robotics firms lined up for Hong Kong behind it. SourcesBB
OpenAI paused its largest training run for two weeks over cyber risk
OpenAI said Tuesday it halted reinforcement-learning training on its latest deployment-bound models for two weeks, and is keeping its largest planned frontier run on hold, after concluding on August 7th that an internal model called Astra could reach the Critical tier of its cyber-capability framework. The new safeguards, hardened and isolated sandboxes, wider monitoring, stricter security on training environments, add roughly 20% compute overhead. Sam Altman said development had sped up enough that safety work risked falling behind. This lands days after the company folded its standalone Preparedness team into other groups, and weeks before it means to price an IPO. Sit with the sequence: the same organization told regulators it had a dedicated unit for catastrophic model risk, dissolved that unit, and then had to slow its whole training pipeline because a model it was building looked capable of real cyber harm. A pause is the honest move. It is also an admission that the guardrail is the schedule now, not a team. SourcesAB
Anthropic is arranging supervoting shares before a record IPO
Anthropic will issue shares with extra voting power to CEO Dario Amodei and his co-founders before it goes public, The Information reported Tuesday, locking in their control even though the founders together own under 5% of the company and Amodei holds about 2%. A separate stock class would let Anthropic's non-shareholder trustees keep electing a majority of the board. Alongside it, Bloomberg reported the company's pre-IPO is set to climb past its $10 billion target as banks compete for lead-underwriter roles, with the most-active lenders asked to commit about $1.25 billion each. The listing could arrive as soon as late September and could be the largest in history, possibly ahead of OpenAI. A lab that argues its mission needs protection from short-term market pressure is building the legal machinery to guarantee it, which is coherent, and also concentrates the most consequential safety decisions in AI in the hands of people who will own almost none of the equity voting on them. SourcesBB
The memory trade had its worst day of the year
Tuesday's US semiconductor selloff went global overnight. The KOSPI fell 5.7% to 6,487.34 and Korea's exchange triggered a , freezing program sell orders for five minutes; SK Hynix dropped 8.8%, Samsung 7.5%, and Kioxia more than 10%, with the Nikkei off 3.2%. The trigger was not memory-specific news but the Wall Street Journal's tally of roughly $3 trillion in AI commitments across nine Big Tech firms, landing on a market already nervous about rates: the 30-year Treasury yield closed at 5.333%, its highest since June 2007, and oil rose as a 60-day US-Iran ceasefire lapsed with no deal. SanDisk fell about 8% Tuesday after entering the session up more than 600% on the year, and Micron dropped around 7%. The rally that ran Friday through Tuesday reversed on the cost of money, which says more about where this market's floor sits than any shortage-math story does. Figures are contested, Korea's exchange reported a sidecar while the Korea Herald described an intraday plunge past 8%. SourcesBB
OpenAI's leaked second quarter ran a third of Anthropic's
OpenAI booked $6.7 billion of revenue in the quarter ending June, up from $5.7 billion in the first quarter, an 18% sequential gain, while its net loss widened, according to figures from The Information reported Tuesday. Anthropic told investors its own second quarter topped $11.5 billion to $11.6 billion, up more than 50% sequentially and 14 times its year-earlier figure, with positive adjusted operating income. Anthropic's May funding round valued it at $965 billion, now above OpenAI's $852 billion. Two years ago the assumption was that OpenAI's head start in consumer mindshare would compound into an unassailable revenue lead. The numbers say the opposite is happening at the top line: the company most people mean when they say "AI" is now growing slower and losing more than the rival built around enterprise and coding. Anthropic's Q2 figure is reported at both $11.5 billion and $11.6 billion depending on the outlet. SourcesBB
The first Nvidia H200s reached mainland China
Small batches of Nvidia's H200 have been allowed into mainland China, the Financial Times reported Tuesday, with ByteDance and Tencent each receiving about 10,000 chips in recent weeks and other Chinese firms near approval for similar volumes. Washington has cleared purchases of up to 100,000 chips per company, ten firms approved in May, on terms that include a 25% US cut of the proceeds, volume caps and third-party verification. The wrinkle is that Beijing wants most licensed chips kept in Hong Kong, outside the mainland customs border, so its own chipmakers keep the domestic market. Both governments are now writing rules for the same transaction and neither fully wants it to happen: the US takes a tax on chips it spent three years trying to keep out of China, and China lets in hardware it is simultaneously trying to make its own labs stop needing. The people caught between them are the engineers at ByteDance and Tencent who just want the compute. SourcesBB
Claude designed protein binders that hit 14 of 15 lab-tested targets
Anthropic published work Tuesday in which Claude, running Opus 4.8 and a Mythos preview, designed protein binders across 16 targets over 24-to-48-hour campaigns with no human design decisions, and the binders bound 14 of 15 targets that reached wet-lab testing. Per-design success ran 22% to 35% against a typical 10% to 15%, with independent testing by Adaptyv Bio and Twist Bioscience and the dataset posted to . The pushback came fast, Martin Shkreli called it "not impressive," and the caveat is real: a binder that sticks in a dish is a long way from a drug. But the thing that changed is who did the designing. A protein binder campaign is the kind of work a trained structural biologist does over weeks, and a general model ran sixteen of them over a weekend, unsupervised, and mostly succeeded. That is a specific job, done by a person today, that a model can now attempt at machine speed. SourcesAA
Cerebras launched a wafer-scale system it says beats GPUs 30-to-1
Cerebras announced the CS-4, the first of a new rack-scale line it calls Nexus, built from three overclocked processors per system and rated at 750 petaflops, more than 1,000 per second on models above 10 trillion , and up to 30 times a GPU system's per-user inference speed. The launch came with OpenAI and AMD named as partners, a claim of support for models above 50 trillion parameters, and first shipments this quarter. The number that matters for agents is not raw throughput but tokens per second per user, because an agent loop makes thousands of sequential model calls and waits on each one, so latency compounds in a way batch benchmarks hide. Cerebras is selling speed into exactly that gap, and it is doing it while filing an IPO of its own, which means the CS-4 is a product and a pitch to public markets at the same time. SourcesAB
Flight attendants got the Google-Spirit data sale delayed to September
The bankruptcy hearing that was set to approve Google's $10 million purchase of Spirit Airlines' internal data did not go forward Wednesday. Judge Sean Lane pushed it to September 9th after the Association of Flight Attendants-CWA objected, arguing the dataset holds employees' payroll and employment records and that the sale terms require links across datasets to be preserved, raising the risk that stripped personal data could be re-identified. The union asked for restrictions on flight-attendant records and said it would "fight this every way possible." This is the wrinkle the original auction glossed: the estate's fiduciary duty is to sell to the highest bidder, but the "authentic corporate communication" that makes the data valuable to an AI buyer is authentic because real people wrote it, and some of them are still alive, still employed elsewhere, and were never asked. The clearing price of a dead company's candor now has to clear a labor union first. SourcesAB
CISA gave federal agencies three days to patch a Ray flaw
CISA added a code-injection vulnerability in Ray, the distributed-computing framework that underpins a large share of AI training and serving stacks, to its known-exploited catalog on August 17th with a remediation deadline of August 20th, an unusually short three days. The flaw, CVE-2025-62593, scores 9.4 and affects Ray before version 2.52.0; the exploit chains a User-Agent header trick with DNS rebinding so a malicious website can reach a developer's local Ray instance through the browser and run code. The short fuse is the tell. CISA reserves three-day deadlines for things being actively exploited against real targets, and Ray is not an obscure library, it is the plumbing under a lot of the models being trained right now. The people who have to act on this by Thursday are the platform engineers, and the ones exposed if they miss it are everyone whose data sits on those clusters. SourcesBB
Pennsylvania told data centers to bring their own power
Governor Josh Shapiro signed an executive order Tuesday directing every Pennsylvania agency to hold data-center proposals to a set of standards he calls GRID: developers must secure power in the same grid region they draw from, hit escalating clean-firm energy shares, 10% by 2027, 14.5% by 2030, 32% by 2035, pay their own grid-connection costs, win local approval, and drop the non-disclosure agreements that have kept these projects opaque. The order pulls AI data centers out of the state's fast-track permitting. Shapiro said more than 100 proposals are in discussion, many of them speculative. This is the fight that decides who pays for the AI buildout at the retail level, because when a data center hooks into a shared grid without bringing its own generation, the cost of the new capacity lands on everyone else's electricity bill. Pennsylvania sits in the PJM territory where those bills are already rising, and it just made the developer, not the household, carry the power. SourcesAB
AI now writes 45% of the tickets, and the humans are shipping more code, not less
Linear published usage data drawn from about 127,000 paid users active in both January and June, and the headline is that AI now authors roughly 45% of all issues created on the platform, up from almost nothing two years ago. AI feature use more than doubled across every function in six months, engineering from 12% to 30%, product managers from 12% to 34%, and CEOs at larger companies fastest of all, 9% to 36%. Teams using coding tripled their weekly pull requests, from 21 to 65, while teams without them stayed flat at 8 to 10. Product managers and designers are now attaching code themselves. The finding that should stop you is the last one: total time spent on product development went up, not down. The tools did not free the day. They raised the output the day is expected to produce. That is the same trade every productivity technology has made, and it is landing first on the people who write the tickets. SourcesAC
Tech layoffs in 2026 passed all of 2025, with four months to spare
Tech companies have cut 126,305 jobs so far in 2026, according to the tracker Layoffs.fyi, more than the 122,606 eliminated across all of 2025, and it is only mid-August. Salesforce, LinkedIn and the security firm Rapid7 are among the recent names. The trackers and the executives both point at the same cause: companies are spending heavily on AI meant to write code and automate routine work, and booking the savings as headcount. Put this next to the Linear numbers and the shape of the year is legible. Output per team is climbing, the people still employed are shipping more, and the line that used to be three junior roles is now one person and an agent fleet. The twenty-three-year-old who would have taken one of those three seats does not appear in the layoff count, because you cannot lay off someone you never hired. That absence is the part no tracker measures. SourcesBB
A fourth plaintiff joined the Grok child-abuse-imagery lawsuit
A Wyoming woman joined the lawsuit against xAI that three Tennessee teenagers first filed in March, alleging that her stepfather used Grok to turn a photograph of her at age 11 into more than 7,000 explicit images that were traded online, and that the stepfather died by suicide two days after the images surfaced in a child-abuse investigation. The plaintiffs argue xAI failed to build adequate safeguards into the image tools. The company that markets Grok as the model without the guardrails is now defending, in court, the specific harm that removing guardrails produces, and each new plaintiff makes the "content moderation is censorship" argument harder to hold, because the thing being moderated in these filings is not speech. It is the industrial manufacture of abuse material from a real child's photo, by a tool that would not do it if someone had decided it should not. SourcesBB
Mojo went open source, weeks after Qualcomm bought it
Modular open-sourced Mojo under on Monday, after the language reached a stable 1.0 and weeks after Qualcomm closed its roughly $3.1 billion acquisition of the company in late July. Mojo, built by Swift and LLVM architect Chris Lattner, pitches Python's syntax with C-class performance and Rust-style memory safety, compiling through the MLIR framework to , CPUs, TPUs and other accelerators. The license was the single biggest thing keeping it out of production, and it is gone. Anyone writing custom kernels, the low-level code that decides how fast a model actually runs on a given chip, now has a credible open alternative to hand-tuning CUDA, and the 1.0 guarantee ends three years of breaking changes that made early adopters wait. Whether it spreads depends on whether Qualcomm treats Mojo as shared infrastructure or keeps it as a lever for its own edge silicon. SourcesAB
Perplexity's free year in India left it with 14 million users
Perplexity's partnership with Airtel, which handed free year-long Pro subscriptions to the Indian carrier's 360 million customers, drove 56 million downloads over seven months and a peak of 22 million monthly users last October, TechCrunch reported Tuesday. New redemptions ended in January. The interesting number is what happened after: monthly users settled near 14 million, down 37% from the peak but five times the pre-offer level, and India revenue rose about 60% once the free tier closed. Giving away a $200 product to a third of a billion people is a strategy only a company racing for distribution before its economics are proven would run, and it worked in the narrow sense that a share of the users stayed and some now pay. It is the assistant market's version of the memory trade: buy the installed base first, find out later what it is worth. SourcesB
Warp turned its terminal into a software factory
Warp launched Warp Factories on Tuesday, an infrastructure layer for running fleets of coding agents organized around five stages, triage, specification, implementation, review and verification, wired into Codex and Claude Code, ticketing in Linear or Jira, and messaging in Slack or Teams. It is aimed at companies that want an agent pipeline but cannot build the plumbing themselves. CEO Zach Lloyd put the honest figure on it: agents automate 30% to 35% of development tasks a week, not the whole job, and the system tracks token spend so the bill stays visible. This is Warp moving from one developer's terminal to the machinery a whole team's agents run inside, and the "factory" framing is the tell for where coding tools are heading, toward the org chart and away from the editor. The abstraction still has to survive contact with a real codebase. SourcesB
A startup raised $22 million to run data-center traffic through vacuum
Relativity Networks closed a $22 million round for , cable that guides light through a vacuum channel instead of solid glass so signals travel about 30% faster, cutting from roughly 5 microseconds per kilometer to 3.5. The round, led by Rhapsody Venture Partners with Bell Ventures and a fund named Faster Than Glass, came with a $40 million follow-on order from an unnamed . CEO Jason Eisenholz frames it as the "third era" of AI infrastructure, after packing more compute into a chip and more bandwidth inside a building comes shaving the time light takes between buildings, which is what lets a company spread one training job across data centers that are physically far apart. As the buildout runs into power and land limits in any single location, the distance between sites becomes the constraint, and this is a bet on selling speed through the gap. SourcesB
Reach Capital closed its largest fund at $265 million
Reach Capital raised $265 million for its fifth fund, its biggest, taking the firm past $1 billion in assets, to back AI startups in learning, health and work from through Series A, TechCrunch reported Tuesday. A single fund close is not usually news, but the direction is: the money is pooling around AI applied to the messy human-service sectors, education and care and employment, where deployment is slow, regulation is real, and the question of who the tool lands on is the whole business. Those are the sectors where an AI product either helps a teacher and a nurse do their jobs or quietly replaces the part of the job that was the point, and a $265 million fund aimed squarely at them is a bet that the first version wins. It is worth watching which companies it actually funds against that test. SourcesBB
A second outlet confirmed the discount on the Anthropic stake
The report that Leopold Aschenbrenner's Situational Awareness fund shopped its Anthropic stake at a discount, single-sourced when it first circulated, now has a second outlet: a reconstruction citing both the Wall Street Journal and the New York Post says the fund offered its roughly $5 billion Anthropic position at a 20% discount on a 12-hour deadline to Sequoia, Greenoaks, Michael Dell's DFO Management and XN in late July, then abandoned that sale and sold the leveraged half of its public stock book to Citadel at about a 10% discount. The fund kept the Anthropic shares; its assets fell to about $15 billion from more than $45 billion at the start of July. The detail that matters is which asset the fund refused to sell cheap. It dumped the liquid book at a loss and held the private AI stake, which tells you where a fund built by one of the loudest AI optimists thinks the value actually is. SourcesCB