It is Thursday, June 26, 2026. OpenAI named its first custom chip after a pepper. Anthropic formally accused China’s largest e-commerce company of running 28.8 million fraudulent conversations with Claude. Google quietly shipped the model it promised six weeks ago and the benchmarks are rewriting the leaderboard. And four senior Google DeepMind researchers have now walked out the door in six days. Here is everything.
OpenAI built its own chip, named it Jalapeño, and used AI to design it faster than any chip in history
On June 24, OpenAI and Broadcom walked into OpenAI’s San Francisco headquarters and physically handed CEO Sam Altman and President Greg Brockman an engineering sample of Jalapeño — OpenAI’s first custom-designed AI inference chip. The ceremony was brief. The implications are not.
Jalapeño was designed from scratch in nine months, from initial design to manufacturing tape-out — a timeline Broadcom CEO Hock Tan described as potentially the fastest high-performance ASIC development cycle in the history of semiconductors. The speed was achieved in part by using OpenAI’s own models to accelerate portions of the chip design process. The same AI being served to users helped design the hardware that will serve future versions of it.
The chip is built specifically for inference — the compute-intensive process of running an AI model in response to a user query — not for training. Every time you send a message to ChatGPT, inference happens. OpenAI currently does that inference on NVIDIA GPUs, which are expensive, power-hungry, and designed for general-purpose computing. Jalapeño is purpose-built around the specific memory movement, networking, and serving patterns of large language models. Early testing shows substantially better performance per watt than current state-of-the-art, though OpenAI has not published final numbers yet. A detailed technical report is coming in the months ahead.
The financial logic matters as much as the engineering. OpenAI burned $3.7 billion in Q1 2026 on $5.7 billion in revenue. A significant portion of that spend is inference compute rented from NVIDIA. If Jalapeño cuts inference costs by even 30 to 40 percent at scale — which is the implied target based on the architecture — the unit economics of ChatGPT change substantially heading into an IPO that needs to show a credible path to profitability. Google has been running its own TPU inference chips for years. Amazon has Trainium. Microsoft has Maia. OpenAI was the last major AI company paying full retail for someone else’s hardware. That era ended on June 24.
Deployment begins before the end of 2026 in prototype scale, with full ramp through 2027 and what Broadcom’s CEO called “full tilt” production in the first half of 2028. The partnership targets gigawatt-scale data centers — compute infrastructure measured in city-level power consumption — alongside Microsoft and other partners.
Read the official announcement: OpenAI’s announcement of Jalapeño | Broadcom’s official press release
Read the coverage: TechCrunch on the chip design and what inference-only means | CNBC with Greg Brockman on the nine-month timeline and AI-assisted design | VentureBeat on the IPO implications and competitive positioning against NVIDIA | SiliconAngle on the Tomahawk networking architecture and rack-level deployment design
Anthropic told the Senate that Alibaba ran 28.8 million fake conversations with Claude to steal its capabilities
On June 10, Anthropic sent a letter to Senate Banking Committee Chairman Tim Scott and Ranking Member Elizabeth Warren accusing Alibaba’s Qwen AI lab of conducting what Anthropic called the largest known distillation attack on its systems to date. The letter was obtained by CNBC and Bloomberg. Alibaba has not responded.
The operation, as Anthropic described it: between April 22 and June 5, 2026, operators affiliated with Alibaba and its Qwen AI lab created approximately 25,000 fraudulent accounts and used them to run 28.8 million exchanges with Claude — specifically targeting the model’s software engineering and agentic reasoning capabilities. The technique is called adversarial distillation. You prompt an advanced model with enormous volume, harvest the responses, and train a cheaper competing model on those outputs. The copying model learns to behave like the original without any access to the original’s weights, architecture, or training data. The research and development cost you bypass is measured in hundreds of millions of dollars.
Anthropic’s letter described the attack as “brazenly” and “illicitly” conducted, and called it a national security issue rather than an intellectual property dispute — specifically because distilled models replicate advanced capabilities while stripping away the safety guardrails and access controls built into the original system. In February, Anthropic disclosed that DeepSeek, Moonshot AI, and MiniMax had conducted similar campaigns involving 150,000, 3.4 million, and 13 million exchanges respectively. Alibaba’s alleged 28.8 million exchange operation is more than double the next largest. Two senators — Bill Hagerty and Andy Kim — are now moving to add an amendment to must-pass defense legislation that would sanction entities conducting such campaigns.
The timing creates a specific irony worth noting. The June 10 letter accused Alibaba of stealing Claude’s capabilities. Two days later, on June 12, the Commerce Department ordered Anthropic to shut down its most capable models globally over export control concerns. Anthropic’s argument that its models need government protection from Chinese extraction and the government’s subsequent action cutting off those models from the entire world happened within 48 hours of each other.
Read the coverage: CNBC with the full letter details and Senate response | Let’s Data Science on the Hagerty-Kim defense amendment and legislative path | Cybersecurity Insiders on what adversarial distillation is and why it is structurally hard to defend against | Android Headlines on the previous DeepSeek, Moonshot, and MiniMax campaigns and how Alibaba’s dwarfs them
Google finally shipped Gemini 2.5 Pro with Deep Think and the benchmarks landed like a dropped hammer
On June 22 — four days before the end of June, and 34 days after Sundar Pichai told the Google I/O audience it would ship “next month” — Google launched Gemini 2.5 Pro with Deep Think. The model is available now in Google AI Studio and the Gemini app for Ultra subscribers. Vertex AI access is rolling out to enterprises.
The benchmark numbers reset the science and reasoning leaderboard. GPQA Diamond — graduate-level physics, chemistry, and biology questions that PhD experts get right roughly 65 percent of the time — came in at 82.4 percent, ahead of Claude Fable 5 at 79.1 percent and GPT-5.5 at 76.3 percent. MMLU-Pro, a broad professional knowledge benchmark, scored 89.8 percent — the highest of any publicly available model. HumanEval+ coding accuracy hit 94.1 percent. Deep Think is Google’s extended reasoning mode: it runs internal chain-of-thought processing before generating output, comparable to Claude Extended Thinking and OpenAI’s o-series. It is specifically designed for hard science, complex mathematics, and multi-step reasoning tasks.
The practical interpretation is more specific than the headline numbers suggest. Gemini 2.5 Pro Deep Think now leads on science, graduate-level reasoning, and hard mathematics. Claude Fable 5 still leads on software engineering and long-horizon agentic coding — the workloads where most enterprise developer spending currently lives. For research teams, life sciences, financial analysis, and complex reasoning, the benchmark shift is meaningful. For teams building coding agents and developer tools, the competitive picture has not changed. The two models are now the best in the world at different things simultaneously.
Pricing: standard mode is estimated at $2.50 per million input tokens. Deep Think mode runs at approximately four times the standard rate, putting complex reasoning sessions in the range of $10 per million input tokens — the same tier as Claude Fable 5’s standard rate.
Read the official announcement: Google DeepMind’s Gemini model page with current benchmarks
Read the coverage: Full launch day breakdown including benchmark context and competitive implications | Build Fast With AI on how Gemini 2.5 Pro stacks up against Fable 5 and GPT-5.5 across different workloads
Four senior Google DeepMind researchers left for Anthropic and OpenAI in six days and the talent drain has a pattern
The most significant talent story in AI this month has not been a single departure — it has been the rate. In the six days between June 18 and June 24, four senior Google DeepMind researchers announced they were leaving, two for OpenAI and two for Anthropic.
The most headline-generating exit: Noam Shazeer, co-author of the 2017 paper “Attention Is All You Need” — the Transformer architecture paper that every AI model in existence today is built on — announced on June 18 he was joining OpenAI as Lead for Architecture Research. Google had paid approximately $2.7 billion in 2024 to bring him back from Character.AI, the chatbot startup he co-founded after leaving Google in 2021. He lasted 22 months. Sam Altman called it a hire he had “wanted since the very beginning of OpenAI.” Alphabet shares closed up 1.17 percent on the news, suggesting investors believe Google’s $422 billion revenue base and compute commitments outweigh any single researcher departure. CNBC’s Jim Cramer called it a coup. The market disagreed with Cramer, which is usually a reasonable indicator.
Two days later, on June 20, John Jumper — the DeepMind researcher who led AlphaFold, the protein structure prediction system that earned a Nobel Prize in 2024 — announced he was joining Anthropic. Dario Amodei said Anthropic’s scientific AI roadmap had “suddenly come into sharp focus” with Jumper’s arrival, and pointed to biology and drug discovery as the area where that focus lands. The competitive context: OpenAI launched GPT-Rosalind in April for biomedical research, and Google’s own Isomorphic Labs spun out of DeepMind to pursue similar drug discovery applications. Jumper leaving DeepMind for Anthropic is that competitive dynamic made personal.
Industry reporting indicates Shazeer and Jumper are part of a broader pattern of departures from Google DeepMind that accelerated in early 2026, driven by a combination of compensation structures, publication restrictions, and the difference between working inside a trillion-dollar advertising company and working at a lab where the primary product is the research itself.
Read the coverage: Build Fast With AI on all four departures and the broader DeepMind talent pattern | Build Fast With AI on John Jumper joining Anthropic and the drug discovery competitive stakes
Also worth reading today
- Anthropic is on track for its first operating profit — approximately $559 million in Q2 2026 at an annualized revenue run rate of $47 billion. The structural reason: Claude Code holds approximately 40 percent of the generative AI coding market, which carries higher margins than consumer chatbot usage. OpenAI projects losses of $14 billion for the full year 2026. The two companies are now on diverging financial trajectories heading into their respective IPOs. (Build Fast With AI June 26 roundup)
- GPT-5.6 is days away. Polymarket prices a June 28 launch at 83 percent probability based on developer tracking of Codex backend logs and OpenAI’s historical release patterns. OpenAI’s Chief Scientist previewed it as a meaningful improvement over GPT-5.5 with a late-June target. No official date has been announced. (Build Fast With AI June 26 roundup)
- The EU AI Act general-purpose AI obligations deadline is four days away. June 30 is the compliance cutoff for any company with a frontier model deployed in Europe. Governance documentation, transparency reports, and systemic risk assessments are due. The list of companies that are not ready is longer than regulators are publicly acknowledging. (Build Fast With AI June 26 roundup)
- Alphabet replaced Verizon on the Dow Jones Industrial Average this week, effective June 23. The addition of a company whose core product is now AI-powered search to the 30-stock index that has tracked American economic strength since 1896 is a useful shorthand for how thoroughly the AI industry has become the center of gravity for the US economy. (CNBC)
That is your Thursday. OpenAI named a chip after a pepper and it might save the company’s margins. Anthropic accused China of running the largest AI theft operation in history and then discovered the government had cut off the model being stolen ten days earlier. Google shipped its best model ever and was immediately outpaced in a different category by the company whose engineers it keeps losing. And the Dow Jones now has an AI company in it. The circus is fully operational. See you tomorrow.