Morning Brief, August 15th 2026
Anthropic's quarterly revenue grew 14-fold and the quarter turned a profit. Alibaba put a frontier-grade 27B model under . OpenAI's executive floor emptied ahead of its IPO. And ads are coming to ChatGPT in Europe.
Anthropic's revenue grew 14-fold, and the quarter turned a profit
Anthropic told prospective investors its preliminary second-quarter revenue passed $11.5 billion, Bloomberg reported Friday, at least 14 times the $787 million it booked in the same quarter last year and more than double the $4.73 billion of the first quarter. The same materials show positive adjusted operating income, the company's first profitable quarter, arriving roughly two years ahead of its own internal projections. The May projections circulated to investors had called for $10.9 billion of revenue and a $559 million operating profit; the preliminary actuals came in above both. The engine is enterprise: coding and security workloads, with compute spend falling from 71 cents per revenue dollar in the first quarter toward 56 cents. Two caveats belong next to the headline. The profit figure is adjusted, a framing critics like Ed Zitron have already gone after, and Anthropic itself has told investors profitability may not hold through the year as scheduled compute costs land. It is still a fact nobody else at the frontier can claim: the lab expected to list in October would go public with a profitable quarter on the books. SourcesBC
The missing Qwen3.8-27B arrived, under Apache 2.0
The 27-billion-parameter sibling that was absent from Qwen3.8-Max's launch appeared on and ModelScope on Friday, and Alibaba shipped it under plain Apache 2.0, the license the 2.4-trillion-parameter flagship conspicuously did not get. The model reads text, images and video natively, carries 262,144 of context extensible to a million, and mixes gated attention with DeltaNet layers in a hybrid stack aimed at 24GB consumer . Alibaba's own numbers put it at 61.7 on SWE-bench Pro, 73.0 on Terminal-Bench 2.1, 89.2 on GPQA Diamond and 84.3 on OSWorld computer use, which would make it the strongest model a single high-end desktop card can run. The two-license split is the tell worth keeping: the frontier model ships under custom terms, and the tier a developer can actually deploy ships free. That is a strategy for making Qwen the default substrate of local and embedded AI while keeping the crown jewels encumbered. No third party has reproduced the benchmarks yet. SourcesAB
OpenAI cleared its executive floor, and Brockman is what remains
A month of departures at OpenAI now reads as a list of the people who used to run it: former COO Brad Lightcap, there since 2018, chief revenue officer Denise Dresser, ethics chief Chloé Bakalar, the head of safety and the chief futurist, with Fidji Simo, Sam Altman's number two, gone since last month. Axios reported Friday that the company frames this as a deliberate pre-IPO refresh, with co-founder Greg Brockman inserting himself at every level to build a leadership team aimed at beating Anthropic in the enterprise. CNBC's read from governance watchers is blunter: a revolving door at this scale, this close to a listing, is a red flag have to price. The numbers argue both ways. Bloomberg reporting has OpenAI on track for $40 billion in annualized recurring revenue, and Brockman told staff July revenue rose 20% month over month with business customers up 32%. Growth like that forgives a lot of churn. What it cannot answer is the question the safety departures raise: nearly everyone whose job was to say no has now left, in the same season the company is optimizing itself for sale. SourcesBB
Apple trained its own model for China, a first for a foreign company
Apple has trained a proprietary for the Chinese market with Alibaba's help, and Beijing's algorithm registry cleared it in July, the first proprietary model from any foreign company the Cyberspace Administration of China has ever approved. The reporting, which surfaced Friday, describes a dual-track strategy: the Apple-trained model runs alongside the Qwen-based integration already approved for Apple Intelligence in China rather than replacing it, with a consumer rollout expected in the coming months. Apple explored Baidu, DeepSeek and ByteDance before settling on Alibaba, whose regulatory standing with the CAC and Qwen optimization for Apple's on-device MLX framework carried the choice. The concession structure is the story. Apple gets a layer of proprietary control no other foreign technology company has achieved in China's generative-AI era, and it gets it by building on a Chinese partner's foundation, inside Chinese rules, against Huawei phones that ship with none of those constraints. Every other Western company wanting AI features in China just watched the price of admission get set. SourcesBB
Free ChatGPT is now unlimited, on a frontier-family model
OpenAI finished rolling out unlimited text chats for ChatGPT's Free and Go tiers this week, with GPT-5.6 Luna, the lightweight member of its flagship family, as the new default. Free users also get a Think button that buys the model more reasoning time per message; caps remain on file uploads, images and tools. OpenAI's internal evaluation says responses containing at least one factual error are about 62% less common with Luna than with GPT-5.5 Instant, the model it replaces, on financial, medical and legal prompts. A billion-plus weekly users now hold unmetered access to a current-generation model, which is a cost decision as much as a product one, and it did not happen out of generosity. The unlimited free tier is the top of a funnel that OpenAI's enterprise push and its coming advertising business both draw from, and giving away the base model is what fills it. The next story is what filling it costs the people inside. SourcesAB
Ads arrive in European ChatGPT this month
OpenAI Ireland notified Free and Go users across the EEA and Switzerland on Friday that advertising will appear in ChatGPT later this month, in an email styled as a privacy-policy update. Ad selection starts contextual only: the current conversation topic, general location and device type. Personalized ads drawing on chat history and memory require explicit opt-in consent under the EU ads policy OpenAI published in June, which leans on consent rather than legitimate interest as its legal basis. Pro, Enterprise, Business and Education tiers stay ad-free, and advertisers see only aggregate views and clicks. This is the first time ads have been switched on inside 's consent regime, after earlier rollouts in the US, UK, Japan and Brazil, and the mechanics matter more than the banner: what fraction of European free users opt in to personalization, and how hard the product nudges them to, will define what an ad-funded assistant is allowed to be everywhere else. The notification arrived one week after the free tier went unlimited. SourcesC
IBM will train tens of thousands of consultants on OpenAI's stack
IBM and OpenAI announced a partnership Thursday that puts GPT-5.6, Codex and ChatGPT Work inside IBM Consulting Advantage, the delivery platform for IBM's global consulting arm, with a dedicated OpenAI practice and plans to train and certify tens of thousands of consultants on OpenAI technologies over the coming months. Target industries are financial services, government, telecommunications and retail, and the deal extends to wiring OpenAI models into IBM Autonomous Security, IBM's agent-based cybersecurity service. Financial terms were not disclosed. The deal lands one day after IBM locked up two years of Together AI's open-model capacity, and the pair of moves describes one strategy: IBM is assembling both frontier-lab and supply, then selling the integration labor, which is the part enterprises actually get stuck on. For OpenAI it is distribution into exactly the regulated industries its new enterprise leadership was hired to win, without building a consulting arm of its own. SourcesAB
Writer says the harness, not the model, sets agent economics
Writer released Palmyra X6 on Thursday alongside research it calls the Harness Effect, and the research is the interesting half. The company rebuilt its orchestration layer and measured a 41% drop in blended cost per task and 44% faster completion across every model it runs, including Claude Sonnet 4.6 and Gemini 3.1, before its new model enters the picture. With Palmyra X6, a 744-billion-parameter mixture-of-experts model post-trained on Z.ai's GLM-5.2 with about 40 billion active , the combined stack runs 52% cheaper and 48% faster with a claimed 10% quality lift. The argument to enterprises is that orchestration design, not model choice, sets the token economics of agentic AI, which if true relocates a lot of pricing power from the labs to whoever owns the harness. Worth noting alongside: a San Francisco enterprise vendor just shipped its flagship on a Chinese open-weight base and said so in the first paragraph, which stopped being a liability disclosure and became a cost story. All figures are Writer's own. SourcesAB
DeepSeek's new prices go live tomorrow, with an IPO behind them
The price schedule DeepSeek published last week takes effect tomorrow at 16:00 UTC, moving V4-Flash and V4-Pro onto peak and off-peak billing with increases of 50% to more than 1,100% depending on model, token type and hour. V4-Pro output goes to $3.96 per million tokens at peak against $0.87 today. The new context is why: Bloomberg reporting ties the increase to a company preparing to go public, with DeepSeek in talks with accounting firms and banks, weighing a funding round at a pre-money near $71 billion, and eyeing a filing as soon as this year for a possible 2027 debut. A lab that built its identity on being the industry's price floor is repricing itself into a business the public market can underwrite, which reframes the increase from a capacity story into a margin story. Whether developers absorb it or route around it is the live experiment, and the off-peak half-price window is the pressure valve. SourcesAB
The AI tape stalled on consumer data, with Nvidia's print 11 days out
US markets pulled back from record highs Friday on weak retail spending and consumer confidence numbers, and chipmakers led the decline, with Broadcom down 6%. The exception stayed exceptional: AMD extended its post-earnings run to close at $514.39, up 130% for the year, on second-quarter data-center revenue that doubled year over year to $6.7 billion, while Nvidia finished at $225.16 with its own report due August 26th. Friday's move had no AI news in it, and that is the read worth taking: nothing touched the story, and the day still showed how little non-AI economy is left holding the index up when the AI names pause. The Korean side of the trade had spent the week making the opposite point, with the KOSPI's memory-driven rally topping 7,000 for the first time. Eleven days out from the quarter that settles the summer, the market is priced for Nvidia to confirm everything its suppliers already said. SourcesBC
Six grid operators owe FERC an answer on data-center power by Monday
The 60-day clock on the Federal Energy Regulatory Commission's June 18th show-cause orders runs out around Monday, and all six US regional grid operators must either prove their existing interconnection tariffs are just and reasonable for loads above 20 megawatts or file rewritten ones. The orders, issued under Section 206 of the Federal Power Act, preliminarily found that current rules fail to address large loads and co-located loads, data centers first among them, across territory serving roughly 200 million people in more than 30 states. The filings due this week will define how fast gigawatt-scale AI campuses can legally connect to the grid in most of the country, which has quietly become the binding constraint on the buildout: the chips ship in quarters, the interconnection queue runs in years. Watch which operators defend their tariffs and which capitulate and rewrite, because the split will map where the next wave of data centers gets sited. SourcesBC
The eval shop the labs cite raised $40 million
Vals AI closed a $40 million Series A led by Andreessen Horowitz at a $400 million valuation, with 8VC, Bloomberg Beta and HRT Ventures participating. Vals runs independent whose results already appear in model cards from OpenAI, Anthropic, Google, Meta and xAI, and its headline finding is the kind of number self-reported never produce: correctly complete under 52% of its finance-analyst tasks. The roadmap is custom benchmarks built from a customer's own GitHub repositories, a cyber benchmark, and an index built with CoreWeave. The raise is the market pricing a structural fact of this week's news: a Chinese lab shipped a model whose headline scores nobody can reproduce for two weeks, and every lab's launch numbers are its own harness configuration. Whoever becomes the auditor of record for model capability holds real power over procurement, and the labs citing Vals in their own cards suggests the position is already half-occupied. a16z partner Jennifer Li made the frame explicit, comparing Vals to Moody's for credit markets. The task-completion figure is Vals' own measurement. SourcesAC
Google is compiling AI to run on encrypted data
Google published an expanded case Friday for HEIR, its open-source compiler that converts trained AI models to run on encrypted inputs using , so the operator of the server never sees the data it is processing. The toolchain compiles a TensorFlow Lite model to encrypted execution, producing a private inference from a three-layer network in about 16 seconds, and Google is pairing it with hardware-acceleration partners Belfort, Niobium, Cornami and Optalysys plus its Jaxite library for running FHE workloads on GPUs and . Named applications are recommenders, fraud detection, intrusion detection and hotword spotting. Sixteen seconds for a three-layer net says exactly how far this is from encrypted LLM inference, and pretending otherwise would be marketing. The direction still matters: the week's news is full of AI systems ingesting data their users would rather not hand over, and a compiler pipeline that makes private inference merely slow instead of impossible is how that eventually changes. SourcesAA
Gemini's visible watermark is now optional
Google will let users remove the visible from AI-generated images, videos and songs, VP Josh Woodward confirmed Friday, via a Media Watermark toggle rolling out in the Gemini app and the Flow video editor for the Nano Banana, Omni and Lyria models. The invisible SynthID signal and metadata stay embedded regardless of the setting, and enterprise administrators can lock the toggle. The timing is the story. Two weeks ago the EU's transparency rules made machine-readable marking of AI content mandatory, and this week a compliance study found nearly half the companies covered by California's detection law had not turned their detection tools on. Google's answer to that landscape is to keep provenance in layers only platforms and regulators can read while letting the human-visible layer go dark. That is defensible engineering and a real transfer of power: the ordinary person looking at an image loses the label, and verification becomes something you need tooling, or a platform's cooperation, to do. SourcesB
Debian is voting on whether AI may write Debian
The Debian project has opened General Resolution 2026-002, a ranked-choice vote among proposals that run from a full ban on LLM-assisted contributions to a permissive framework with disclosure requirements. The ban proposal, led by Matthias Geiger, rests on copyright uncertainty, quality control and objections to scraped training data, and would cover source code, web resources and official communications. The permissive alternatives require contributors to fully understand and be able to explain anything they submit, take responsibility for licensing, and flag AI assistance in commit metadata. Discussion opened July 24th, ballots go to every Debian developer, and the outcome will set the project's stance for years. This is the most consequential open-source governance vote on AI yet held: Debian's decision propagates into the derivatives and the norms of half the Linux ecosystem, and it is being made by the maintainers who carry the liability, not the vendors selling the tools. SourcesAB
Grok 4.6 reached GitHub Copilot two days after release
xAI's Grok 4.6 began rolling out in GitHub Copilot on Friday, selectable across eight surfaces including VS Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, the and the cloud agent, for paid Copilot tiers with an admin opt-in for business and enterprise plans. The model shipped August 12th; the gap between a frontier release and its arrival on the largest developer distribution channel is now measured in days, and that velocity is the story more than the model. Copilot's model picker has become the app store of coding models, with Microsoft distributing whatever wins benchmarks regardless of whose stack it competes with: Grok now sits one dropdown away from OpenAI's and Anthropic's models inside a Microsoft product, days after SpaceX closed its purchase of Cursor to compete with all of them. Positioning is xAI's, and GitHub's internal claims about long-horizon task performance are not independently benchmarked. SourcesAB
Forty companies became unicorns in July, the most in four years
Crunchbase's Unicorn Board added 40 companies in July, the highest monthly count in more than four years, with three entering above $10 billion and the board adding over $100 billion in value for the second straight month. The sector list is the AI economy in miniature: financial services, robotics, AI orchestration, AI, energy and semiconductors. Geography split 19 US, 8 China, 3 UK, 2 Singapore. A monthly cohort this size has not appeared since the 2021 ZIRP vintage, and that comparison cuts both ways: it is what a genuine platform shift looks like from the funding side, and it is also what the top of a cycle looks like, and the two are indistinguishable from inside the month they happen. The difference this time is concentration: 2021 spread its cohorts wide, and this one puts fewer, larger bets into capital-hungry categories, which is either discipline or just a bigger blast radius. SourcesA
SMIC may bolt more tools into fabs it already has
SMIC is weighing additional equipment at existing sites after AI-adjacent demand outran its forecasts, co-CEO Zhao Haijun said, with future wafer starts "far exceeding our previous expectations" and customer orders visible through the end of 2027. The demand is not the leading-edge parts target: it is the mature-node chips that ride alongside AI processors, power-management BCD parts for servers and data centers first among them, where shortages are spreading. This adds the capacity leg to the pricing leg from Friday's earnings, where SMIC posted its first quarter above $3 billion and said it would raise wafer prices again in the third quarter. China's largest is now capacity-constrained on the unglamorous half of the AI , the half no export control reaches, and it is now expanding into that demand on top of repricing it. Details are promised in future briefings. SourcesBB
The most-read new paper is plumbing for choosing which model to call
Topping Hugging Face's trending papers is LLMRouter, from UIUC, an open-source infrastructure for building, evaluating and deploying model routers, the layer that decides per query which model to call under quality and cost constraints. The paper formalizes routing as a sequential decision process, ships a benchmark, xRouteBench, spanning generic, memory-augmented, vision, time-series and personalized routing, and packages the whole pipeline from training data construction to deployment. The timing explains the readership. In one week DeepSeek tripled peak pricing, Google halved Gemini 3.7 Flash's launch price, OpenAI made a frontier-family model free without limits, and Writer published research saying the harness sets the economics. When prices diverge that fast across near-substitute models, per-query routing stops being an optimization and becomes the cost structure, and open tooling for it is infrastructure the whole application layer was missing. It is a preprint, and router quality claims await outside replication. SourcesAB
Editorial
The free tier is the product now
Look at what happened to the price of intelligence this week, from the two ends at once. At the top, DeepSeek, the company that set the industry's floor, raises prices up to 1,100% tomorrow. Anthropic books the frontier's first profitable quarter on enterprises paying rack rate. Memory that AI servers need costs double what it did a generation ago. At the bottom, OpenAI hands more than a billion people unlimited access to a current-generation model for nothing, and one week later mails Europe a notice that the ads start this month.
These are not opposite moves. They are one move, seen from two altitudes. The organizations that can pay for intelligence are being repriced upward, because they have revealed they will pay. The people who cannot are being aggregated, because they are the inventory. A free user asking Luna about a medical bill or a rental dispute is not the customer in that transaction; the advertiser who wants to appear next to that conversation is. Every ad-funded medium in history has run on this geometry, and there is no reason to think the most personal medium ever built will be the exception.
Here is who it lands on. The person on free ChatGPT in Berlin or Warsaw gets an email dressed as a privacy update, and later this month the machine they ask about debt, health and visa problems starts carrying placements selected by the topic of the conversation. Contextual only, for now, and the personalization is opt-in, because GDPR forced it to be. The person in Ohio got no such email, because nothing forced one. The enterprise buyer, meanwhile, gets a rate card, an SLA and a certified consultant from IBM, because people who pay get contracts and people who don't get policies.
The optimists' case is real and I want to state it fairly: unlimited frontier-family AI at zero price is the largest transfer of capability to ordinary people in the history of this industry. A student in Lagos and a paralegal in Manila now hold a tool that outperforms what a Fortune 500 analyst had eighteen months ago, and they hold it free. If the ads stay contextual, that trade is good, full stop.
I don't believe they stay contextual. The pattern from search and social is that consent screens soften, defaults migrate, and the opt-in becomes a ritual most people click through, because the product is genuinely useful and the cost is invisible. The number that will tell us is the EEA personalization opt-in rate. If OpenAI publishes it, watch it. If OpenAI never publishes it, that is the answer.
What would prove me wrong is specific: ads still contextual-only in the EEA by mid-2027, the free tier still on the flagship family, and no default-nudged personalization prompt in the login flow. Anthropic is running the control group, profitable on enterprise revenue with no consumer ad business at all. Two models of who pays for intelligence are now live against each other, and for once the experiment has a readout.
Nour Haddad
Prediction Watch
Where today's news leaves our open calls. Each one links to the full prediction, its reasoning and the exact test that settles it.
New call: Apple's own China model reaches Chinese users this year (Prediction 2026-08-15-T1). Apple trained a proprietary model for China with Alibaba's help and won the first foreign clearance from Beijing's algorithm registry, with reporting pointing to a rollout in the coming months. We put it at 0.70 that Apple Intelligence features running on Apple's own China-registered model are generally available in mainland China by December 31st. Regulatory clearance is done; the risk left is execution and Apple's own caution. Settles December 31st 2026.
Supporting evidence: Anthropic IPOs before OpenAI (Prediction 2026-08-06-F1). We said Anthropic's shares trade publicly before OpenAI's do. Anthropic just showed prospective investors a 14-fold revenue jump and its first profitable quarter ahead of an expected October listing, while OpenAI shed its COO, CRO, ethics and safety chiefs in a month and is reported to lean toward 2027. One company is polishing an ; the other is still rebuilding the org chart that would sign one. Settles August 6th 2027.
No change: DeepSeek holds the August 16th price increase (Prediction 2026-08-13-B1). We said the up-to-1,100% increase survives 90 days, at 0.70. It takes effect tomorrow at 16:00 UTC, and Bloomberg's reporting that DeepSeek is preparing an IPO strengthens the motive to hold it without yet testing whether it holds. Settles November 14th 2026.
No change: a Chinese lab open-weights a model at 2.8T parameters or larger (Prediction T5). We said someone in China publishes downloadable at or above Kimi K3's scale. This week's open-weight energy ran the other direction: Alibaba's Apache 2.0 release is a 27B built for consumer GPUs, and the largest open release on record stays 2.4 trillion parameters. Settles February 28th 2027.
China and open weights: the substrate strategy, in three moves. Alibaba put Qwen3.8-27B under Apache 2.0, Apple built its China model on Alibaba's foundation, and Writer shipped its American enterprise flagship on Z.ai's GLM-5.2. Chinese open weights are becoming the base layer other companies build on, on both sides of the Pacific. Z.ai's GLM-5.3 weights remain gated for roughly two more weeks, and DeepSeek reprices tomorrow.
What did not happen. Nothing settled today. Unitree has not yet traded; its debut window opens Monday. Z.ai's GLM-5.3 weights have not been released, so its benchmark claims remain unverifiable. No enforcement action has named a company since obligations went live August 2nd. And no US grid operator filed its FERC response early.
Sources
- Anthropic Revenue Ahead of IPO Surges Over 14-Fold in Second Quarter — Bloomberg B
- Anthropic reports 14-fold revenue increase in Q2 — Crypto Briefing C
- Anthropic's "Profitability" Swindle — Where's Your Ed At C
- Qwen3.8-27B — Hugging Face A
- Alibaba's Qwen team releases Qwen 3.8 models with open weights under the Apache 2.0 license — The Decoder B
- OpenAI sheds senior execs in pre-IPO refresh — Axios B
- OpenAI talent exodus raises 'huge red flag' ahead of IPO — CNBC B
- OpenAI Lost Two Execs in a Week. At Least Revenue Hit $40 Billion — Yahoo Finance B
- Apple Trained Own AI Model for China Market With Help From Alibaba — MacRumors B
- Apple trained its own AI model for China, and handed the brain to Alibaba — The Next Web B
- Improving GPT-5.6 Sol in ChatGPT, and expanding access to GPT-5.6 Luna for free users — OpenAI A
- OpenAI announces unlimited GPT-5.6 Luna access for ChatGPT free users — Neowin B
- ChatGPT Free and Go users in Europe face ads from later this month — PPC Land C
- IBM Partners with OpenAI to Accelerate Secure AI Deployment for Enterprises — IBM A
- IBM partners with OpenAI to bolster enterprise AI push — TechCrunch B
- WRITER Makes Agentic AI Economically Sustainable at Enterprise Scale With Palmyra X6 — Writer A
- Writer says its new Palmyra X6 model cuts AI agent costs by 52% — VentureBeat B
- DeepSeek API pricing — DeepSeek docs A
- DeepSeek Readies IPO Filing in China, Weighs New Funding Round — Bloomberg B
- Stock Market Today, Aug. 14: Rally Stalls on Disappointing Consumer Data — Yahoo Finance B
- AMD and Ouster rise: top AI stocks today, August 14, 2026 — StartupHub C
- What data center developers need to know about FERC's large load directives — Utility Dive B
- FERC Presses Grid Operators on Data Center, Large Load Interconnections — Morgan Lewis C
- Series A: Always a Higher Peak — Vals AI A
- Vals AI Raises $40M From a16z: Frontier Models Fail 52% of Real Finance Analyst Tasks — Tech Times C
- How Google is Making Private AI Practical with Homomorphic Encryption — Google A
- google/heir — GitHub A
- Google will now allow users to remove visible watermark from its AI generations — TechCrunch B
- General Resolution: LLM usage in Debian — Debian A
- Debian project faces fundamental decision on LLM use — heise online B
- Grok 4.6 is now available in GitHub Copilot — GitHub Changelog A
- Grok 4.6 Arrives in GitHub Copilot Across Eight Development Surfaces — Unite.AI B
- 40 Companies Joined The Unicorn Board In July, The Highest Count In 4 Years — Crunchbase News A
- SMIC weighs more capacity as AI-related chip demand exceeds forecasts — South China Morning Post B
- China's SMIC Eyes Surprise Capacity Boost on AI Chip Demand — Seoul Economic Daily B
- Meta Patents AI Glasses to Use Facial Recognition to Identify People, Make Highlight Reels of Your Dinner Party — 404 Media B
- LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers — arXiv A
- Trending papers — Hugging Face B