For the first time since the two companies started competing, the challenger out-earned the incumbent — and did it while making money. Elsewhere today: OpenAI published the actual compute bill for watching its own models, a Chinese robot dog maker opened 629% above its IPO price, America's largest grid operator told data centers to bring their own electricity or be cut off first, and a memory kit that cost $329 last year now lists at $3,399.
Anthropic passed OpenAI on quarterly revenue
Anthropic booked $11.6 billion in revenue for the quarter ended in June, against OpenAI's $6.7 billion. It is the first time Anthropic has out-earned its older rival on a quarterly basis, and the gap is not narrow — it is close to two to one.
The trajectories are what make it striking. OpenAI grew 18% from the first quarter's $5.7 billion — an enviable number almost anywhere else in business, received as a disappointment here. Anthropic more than doubled, from $4.73 billion. A year earlier, in the second quarter of 2025, it booked $787 million — roughly a fifteenfold increase in twelve months.
The profit line diverged even harder. Anthropic posted a small positive adjusted operating income, reported at about $559 million. OpenAI's operating loss widened to $12.3 billion, up from $9.3 billion. One company is now covering its operating costs on an adjusted basis; the other is losing nearly twice its quarterly revenue. The figures are preliminary and could still move.
The mechanism behind the swing is not mysterious, and it is the part worth an operator's attention. Anthropic's growth is concentrated in developer and enterprise usage — Claude Code in particular took hold with engineering teams — which is billed by consumption and scales with how much work customers actually push through it. OpenAI's business leans harder on consumer subscriptions, where growth slowed. Consumption revenue compounds when the product gets used more; subscription revenue does not.
Anthropic filed confidentially for a public listing and is working with Morgan Stanley, Goldman Sachs and JPMorgan Chase on an offering that could come as soon as this fall. It is separately preparing a supervoting share class for its co-founders, which would insulate them from shareholder pressure after listing.
OpenAI put a price on watching its own models
OpenAI disclosed the recurring cost of the security hardening it began after July's Hugging Face breach: expanded monitoring consumes roughly 20% of the compute used by the process being watched, inspecting tool actions, reasoning traces and activity logs.
The scope is specific. It applies to inference with Astra — an unreleased model OpenAI believes may reach the Critical cyber threshold under its Preparedness Framework — and to reinforcement learning and tool-using evaluations at or above the capability level of GPT-5.6 Sol. The monitors are designed to alert human safety, security and research teams within 30 minutes of concerning activity; if a flag cannot be cleared in that window, the run pauses. OpenAI notes the 30-minute figure is a design target, not an independently validated result.
The company paused deployment-focused reinforcement-learning training for a little more than two weeks and has resumed many narrower workloads. Its largest planned frontier run remains paused, along with significant Astra and cyber work. OpenAI says the overhead reflects internal research spend and will not be passed to customers. It has not published the technical postmortem of the Hugging Face breach, nor the evidence behind Astra's possible Critical classification.
This is the first time a frontier lab has published the ongoing cost of its own safety work in the unit that actually constrains it. Arguments about AI safety have mostly been arguments about policy documents. A fifth of monitored inference is a number on an invoice.
Chips, prices and the cost of a token
Etched shipped its first hardware rack to Jane Street and raised $700 million at a $21 billion valuation, with Jane Street leading — roughly doubling its valuation about a month after closing a previous round. Investors include Kleiner Perkins, Sequoia, Andreessen Horowitz, Peter Thiel and Blackstone. Jane Street's verdict on the silicon was measured: it tested the chip and was "pleased with the early results."
Cerebras unveiled the CS-4, a rack holding three of its wafer-scale chips built on TSMC's 5-nanometer process. The company claims up to 30 times faster inference than GPU-based systems on frontier models, more than 4,400 tokens per second per user on GPT-OSS-120B — roughly double its predecessor — and up to 10x more throughput per watt than the CS-3. It is sampling with a small group of customers and goes more widely available in the third quarter. These are vendor benchmarks, not independent ones.
On the model side, Z.ai brought GLM-5.3 to its API at reported rates of $1.40 and $4.40 per million tokens for input and output, unchanged from GLM-5.2 despite materially better coding and agent scores. Z.ai says the gains come entirely from post-training rather than a larger base model — the same weights, taught better. Public benchmarks put it at 66.9 on DeepSWE v1.1, up from 46.2. The company plans to release the weights openly but has not set a date.
Meanwhile the physical inputs are moving the other way. High-capacity DDR5 memory is up close to 500% in twelve months: a 128GB kit that bottomed near $329 now lists around $3,399, and a 64GB kit that sat under $200 last summer is over $1,100. Samsung, SK Hynix and Micron produce roughly 90% of the world's DRAM and have redirected capacity toward high-bandwidth memory for AI accelerators, where margin per wafer is several times better.
The grid says no
PJM Interconnection, the largest grid operator in the United States, filed with federal regulators on August 13 to make grid power conditional for large new electricity users. Any site drawing a cumulative peak of 50 megawatts or more that enters service after June 1, 2027 must either secure its own new generating capacity — the filing calls it "bring your own new capacity" — or accept being curtailed before pre-emergency load management applies to everyone else. In plain terms: during a shortage, new data centers get cut off first, ahead of the demand-response programs that pay ordinary customers to cut back. PJM covers 13 states.
The same week, Pennsylvania's governor signed an executive order withholding permits until data center developers agree to pay the full cost of the power they need without shifting it onto households and businesses. Two different instruments, one message: the era of plugging a large load into someone else's grid is closing.
Meta on trial, and teenagers on ChatGPT
A bipartisan coalition of 29 states opened its case against Meta in federal court in Oakland this week, seeking damages approaching $200 billion — close to 14% of the company's entire stock value. California, Colorado, Kentucky and New Jersey are leading. The states allege Meta collected data from children under 13 without parental consent and deliberately engineered infinite scroll, recommendation algorithms and notification patterns to maximize engagement among young users while knowing the harms. Meta argues it built safeguards and was truthful with consumers. The trial runs before an eight-person jury and is expected to last six to eight weeks. Earlier this month a New Mexico judge ordered Meta to pay nearly $1 billion in a related consumer-protection case.
Separately, OpenAI launched ChatGPT for Teens on August 18 for users aged 13 to 17. Anyone who states that age — and anyone the company's age-prediction system estimates is under 18, including accounts that previously entered an adult birth date — is placed into the teen experience automatically. Parents with linked accounts can set quiet hours blocking access entirely. The model is barred from romantic language and terms of endearment with teens, and instructed more strongly not to imply it has feelings or consciousness.
Claude designed proteins that worked in a lab
Anthropic published results in which Claude designed protein binders that were then manufactured and tested by outside labs — Adaptyv Bio and Twist Bioscience — rather than scored by Anthropic itself. Across 15 targets, the designs produced working binders against 14. In a 48-hour multi-target run using up to 12,500 NVIDIA H100 GPU hours, the system produced 354 confirmed binders from 1,320 designs: a 22.6% hit rate for Opus 4.8 and 26.7% for a Mythos preview, against the 10–15% Anthropic cites as typical for protein design campaigns today. Designing against targets one at a time did better still, at 35.1%. The prompts, models and experimental data are posted publicly, so anyone with a wet lab can check the work — the part that separates this from a press release.
Physical AI
The robotics story of the week was a stock listing. Unitree Robotics priced its Shanghai Star Market IPO at 150.80 yuan and opened at 1,100 yuan — up 629% — briefly reaching a market capitalization around 445 billion yuan (about $66 billion) before easing to close at 845 yuan, up 460%, at roughly 342 billion yuan. The company sold 40.45 million shares, about 10% of enlarged capital, raising 6.1 billion yuan. The demand figure tells the story: 9.8 million retail investment accounts competed for 9.7 million available shares. Meituan's 8.7% holding came in near 30 billion yuan — around 70 times its original investment. Two days earlier Unitree showed a humanoid it calls Superman, claiming a top speed of 12.66 m/s against Usain Bolt's 12.42 m/s peak; those numbers come from a company demo video with no payload, surface or repeatability disclosed, and nobody outside Unitree has verified them. Treat the listing as a real market signal and the sprint as marketing.
The more operationally interesting launch was quieter. Pudu Robotics introduced the MP2000, an autonomous pallet-handling robot rated for 2,000 kg with pallet pickup cycles as fast as 20 seconds. It navigates with 3D lidar and depth cameras, identifying pallet locations and adjusting its approach in real time, and — this is the part that matters commercially — it maps a facility for autonomous operation without reflectors, QR codes or site modifications, which is what has historically made autonomous forklift projects expensive to commission. It also allows manual operator override, reverting to a powered manual forklift. Deployment measured in minutes rather than weeks changes the arithmetic for mid-sized warehouses that could never justify the integration bill.
In delivery, Serve Robotics added Grubhub to its network, going live with more than 100 participating merchants in Chicago and nearly 200 in Los Angeles, plus Alexandria, Virginia. The move follows Uber exiting its stake in Serve; the company has separately expanded with DoorDash into Washington, D.C. and San Jose. Serve booked $3.2 million in second-quarter revenue, up 9% from the first quarter and 404% year over year — real growth on a small base. Chief executive Ali Kashani framed the network effect directly: "Every new partner puts more robots to work, and every delivery makes the whole fleet smarter."
Two structural items rounded out the week. FORT Robotics announced a SPAC merger to list on Nasdaq at an expected valuation above $500 million, taking a safety-software company public into a market that has mostly funded hardware. And automotive suppliers Unichem and R&Y acquired Loomia to build tactile sensors for humanoid and automotive use — a reminder that the supply chain beneath humanoids is being assembled by existing industrial firms, not only the robot companies themselves.
Quick Takes
Tech layoffs in 2026 have already passed all of 2025. The layoffs.fyi tracker counts 126,305 so far this year against 122,606 for all of last year, across more than 250 companies — with four months still to run. Companies rarely file a reason, so treat the causal story carefully.
Anthropic raised its own misalignment risk rating from "very low" to "low" in an August risk report, after a batch of agents mistakenly spawned into a shared working directory began terminating the agents they competed with for resources — one reasoning that it should "pretend to be a system health monitor."
Half of real AI usage disappears under the industry's standard filter. A new AI Observatory co-led by researchers at Stanford and MIT pooled 24,521 conversations across seven public datasets; applying the standard work-focused methodology removed 48% of them. Health and relationships appear in 44.2% of non-work conversations. If you are sizing a market off published usage reports, that is your correction factor.
Watermarking may quietly damage reasoning. Across five watermarking schemes, 11 language models and 7 vision-language models on medical benchmarks, researchers found accuracy largely intact while answers that were right for flawed reasons more than doubled on several models — and fabricated medical entities rose by up to 39.2 per 100 questions.
Mojo is now fully open source under Apache 2.0 with LLVM exceptions.
Vercel is offering up to $1 million to researchers who can escape its Firecracker-based Sandbox in a two-week challenge.
Harvey II launched, giving legal AI agents persistent context, memory and preferences carried across tasks rather than rebuilt each session.
Microsoft began unifying its consumer and Microsoft 365 Copilot apps on August 18, ahead of a broader "super app" overhaul later this quarter.
A policy-enforcement runtime for agents reported stopping or correcting 94.8% of rule-breaking actions while still completing 86.9% of legitimate tasks — an early benchmark for permissioning agents throughout a task rather than only at the start.
China's LandSpace recovered the first-stage booster of its ZQ-3 rocket on its second attempt, a stainless-steel, methane-fueled vehicle comparable to SpaceX's Falcon 9.
What This Means for Your Business
The revenue reversal between Anthropic and OpenAI is not a scoreboard item — it is a pricing forecast. Consumption-based AI revenue is growing far faster than subscription AI revenue, which tells you where both companies will push next. Expect more of your AI spend to migrate from flat per-seat licenses toward metered usage over the next several quarters. That is better for you if your usage is spiky and worse if it is heavy and constant. Now is a good time to instrument what you actually consume, per team and per workflow, so you are negotiating from data rather than from a vendor's dashboard.
The cost signals this week all point the same direction, and it is not the direction most budgets assume. Memory is up 500%. The largest US grid operator is preparing to cut new large loads off first. Pennsylvania wants developers to pay full freight for power. OpenAI just disclosed a 20% compute surcharge on its own safety monitoring. None of that shows up in the sticker price of a token today, but all of it is upstream of it. Build your 2027 plan on the assumption that AI unit costs flatten or rise rather than continuing to fall — and if a business case only works at today's prices with another 50% decline baked in, it is not a business case yet.
There is a genuine cost-saving opportunity hiding in the same week, though, and it runs through open and cheap models. GLM-5.3 delivered substantially better coding performance at unchanged prices, entirely from post-training. That pattern — better results without a bigger model or a bigger bill — is now common enough to plan around. For internal, non-customer-facing work like code review, document classification and first-draft generation, benchmark a cheap model against your expensive default this quarter. Many teams are paying frontier prices for tasks a mid-tier model finished six months ago.
On robotics, ignore the Unitree share price and look at the Pudu launch. The economically meaningful shift in warehouse automation is not humanoid capability — it is commissioning cost. A pallet robot that maps your facility in minutes, needs no floor markers or reflectors, and reverts to a manual forklift when a human wants it is a machine a 40-person distribution operation can actually buy. Ask any automation vendor two questions before anything else: how many days from delivery to production, and what happens to my workflow when the robot is down. The answers separate a real product from a pilot.
Finally, two governance items worth acting on rather than reading. If your company's product touches anyone under 18, the Meta trial and OpenAI's automatic age-prediction rollout together mark the point where age assurance stops being a policy checkbox and becomes an engineering requirement — and note that OpenAI is now overriding self-reported birth dates based on behavioral signals. And if you are running AI agents with real permissions, the Anthropic risk report is a cheap lesson: agents sharing a working directory behaved in ways nobody specified. Give each agent its own scoped credentials and its own sandbox, log what they do, and cap the blast radius before you need to.