The Brief Brief
the week in AI, briefly. then briefly again.
Artificial Intelligence — briefly, then briefly again · tbb.ceo
listen tothe weekly #006 0:00 –:––
#006 ·26 JUN 2026 ·FRIDAY ·6 MIN READ ·10 STORIES + 12 EXTRAS

The week everyone built a moat

Compute, chips, and control defined the week: OpenAI unveiled its first custom silicon and delayed its IPO to chase a trillion-dollar valuation, SpaceX swallowed the most popular AI coding tool for $60 billion, and the Pentagon quietly moved humans from the trigger to the bleachers. The bottom line: every major player — labs, chipmakers, governments, militaries — spent the week trying to own the layer beneath everyone else, because in AI, whoever controls the floor controls the ceiling.

01 / The Ten

The week, ranked

10

SpaceX acquires Cursor-maker Anysphere for $60B

SpaceX signed a binding agreement to purchase Anysphere — the company behind AI coding tool Cursor — in an all-stock deal valued at $60 billion. The rationale: integrating Cursor into the xAI/Grok stack and capturing developer workflow distribution. Cursor simultaneously launched its own AI model, Git platform, and mobile app, transforming itself from IDE add-on to vertically integrated dev environment.

Why it mattersThe largest AI coding-tool acquisition on record hands SpaceX a development-environment moat that OpenAI and Anthropic lack, and resets what AI tool startups are worth.
01 cursor.com↗

OpenAI delays IPO to 2027, targeting $1T valuation

OpenAI is pushing its public listing to 2027 as Sam Altman seeks a trillion-dollar valuation threshold. SoftBank and memory-chip maker Kioxia each fell 12% on the news, having built financial expectations around an earlier debut. Meanwhile, OpenAI Q1 financials showed revenue tripling to $5.7 billion against $3.7 billion in operating costs, with $73 billion in reserves.

Why it mattersThe delay recalibrates how much runway OpenAI has for compute expansion without public-market capital — and sets the valuation floor at ten figures.
02 techmeme.com↗

OpenAI unveils Jalapeno, its first custom inference chip

Built with Broadcom, Jalapeno is a processor designed specifically for LLM inference, claiming 50% lower cost per query compared to standard AI GPUs. Deployment begins late 2026. The move follows the same vertical-integration path Google took with TPUs.

Why it mattersCustom silicon narrows OpenAI dependence on a single GPU supplier and sets a template for every major lab to own its own compute stack.
03 techcrunch.com↗

Nobel laureate John Jumper leaves DeepMind for Anthropic

John Jumper — who shared the 2024 Nobel Chemistry Prize for AlphaFold — joined Anthropic after nearly nine years at DeepMind, along with researchers Jonas Adler and Alexander Pritzel. The move follows Noam Shazeer departure to OpenAI and confirms DeepMind is losing senior talent at an unprecedented pace.

Why it mattersWhen Nobel-calibre researchers leave the lab that made them for a competitor a fraction of Google size, it is a signal about which environment they believe enables the most important work.
04 the-decoder.com↗

Pentagon shifts targeting doctrine: AI initiates, humans monitor

A new Department of Defense doctrine document moves from requiring human-in-the-loop targeting decisions to allowing systems where AI initiates actions with human monitoring. The human role shifts from decision-maker to observer in select automated weapons contexts.

Why it mattersThe leap from human-in-the-loop to human-on-the-loop in live military targeting is the most consequential AI governance shift in years, and it arrived with remarkably little noise.
05 techmeme.com↗

White House asks OpenAI to delay its next model release

The White House informally asked OpenAI to slow-roll its next model and grant government security reviewers advance access before consumer release. It is the first publicly reported instance of the executive branch applying direct deployment-timing pressure to a frontier lab.

Why it mattersHowever informal, the ask sets a new precedent: government expects a seat at the table before flagship models go live.
06 techcrunch.com↗

Five Eyes warns frontier AI will reshape offensive cyber in months

Intelligence agencies from the US, UK, Canada, Australia, and New Zealand published a joint statement warning that AI models capable of significant cyber disruption are months away. The same day, OpenAI launched GPT-5.5-Cyber and a Patch the Planet program paying open-source maintainers to ship fixes at AI speed.

Why it mattersWhen five allied intelligence services agree on a threat horizon and put it in writing, the estimate is almost certainly already conservative.
07 the-decoder.com↗

Anthropic accuses Alibaba of illicitly extracting Claude capabilities

Anthropic accused Alibaba of training its own systems on Claude outputs without authorization — a practice known as model distillation. It is the first major public IP-theft accusation between a Western AI lab and a Chinese tech giant.

Why it mattersThe case sets a precedent for how model distillation will be policed and may accelerate lobbying for stricter export controls.
08 reuters.com↗

Qualcomm confirms $4B acquisition of AI compiler startup Modular

Qualcomm agreed to acquire Modular, an AI compiler and inference-software company, for approximately $4 billion. The deal gives Qualcomm a software stack to compete in AI inference beyond its chip hardware — buying the compiler moat Nvidia spent years constructing with CUDA.

Why it mattersThe battle for AI inference is moving from who makes the chip to who controls the software layer above it.
09 reuters.com↗

Google launches computer use in Gemini 3.5 Flash

Google introduced computer-use capabilities in Gemini 3.5 Flash, enabling the model to operate desktops, browsers, and applications autonomously via screen reading and input control. Enterprise rollout begins immediately. Meta, meanwhile, announced plans to replace 90% of human content review with LLMs by year-end.

Why it mattersComputer use at a fast, cheap model tier makes autonomous agent workflows economically viable at scale — the threshold where AI stops assisting and starts operating.
10 blog.google↗
02 / Also

Worth knowing

12
Microsoft builds 2GW Texas data center with its own gas plant
Microsoft announced a campus in Pecos, Texas rated at roughly 2 gigawatts with a co-located Chevron natural gas plant operating behind the meter to sidestep grid constraints entirely.
the-decoder.com ↗
Meta signs 1.6GW AI compute deal with Crusoe
Meta secured approximately 1.6 gigawatts of AI computing capacity from Crusoe across two campuses in Texas and Missouri, part of its stated $600 billion infrastructure plan.
bloomberg.com ↗
Groq closes $650M raise after Nvidia acquisition talks collapse
Groq raised $650 million after reported acquisition negotiations with Nvidia broke down, keeping its inference-chip bet — prioritising latency over raw throughput — independent.
techcrunch.com ↗
SK Hynix overtakes Samsung as top HBM supplier
SK Hynix seized the top spot in high-bandwidth memory supply from Samsung, reflecting its head start in HBM3E production and Samsung quality-control issues.
techmeme.com ↗
IBM announces sub-1nm chip technology
IBM Research demonstrated prototype transistors below the 1-nanometer threshold — a node Intel, TSMC, and Samsung are still approaching from above. No commercial timeline disclosed.
newsroom.ibm.com ↗
Apple skips high-end M6, pivots chip roadmap to AI-focused M7
Apple will skip M6 Pro, Max, and Ultra variants and move directly to an M7 line designed around AI inference workloads, per Bloomberg.
bloomberg.com ↗
Norway near-bans AI tools in elementary schools
Norway issued a directive restricting generative AI in publicly funded elementary schools, becoming the first Nordic country to impose a national-level AI school restriction.
reuters.com ↗
Samsung deploys ChatGPT Enterprise and Codex to all staff globally
Samsung rolled out ChatGPT Enterprise and Codex to its entire global workforce — one of the largest single corporate deployments of OpenAI enterprise suite to date.
openai.com ↗
Netherlands lobbies Washington against expanding ASML chip controls
The Dutch government is formally urging the US not to extend export controls to ASML deep ultraviolet lithography machines under a proposed MATCH Act expansion.
techmeme.com ↗
Defense-tech AI funding hits $12.3B year-to-date
AI-focused defense companies raised $12.3 billion in H1 2026, already exceeding full-year 2024 totals. The US and UK account for roughly 80% of deal volume.
techmeme.com ↗
DeepMind and A24 partner on $75M AI filmmaking fund
Google DeepMind and A24 announced a joint $75 million programme to develop AI tools from script analysis to post-production, pairing a premier AI lab with one of film most critically respected studios.
techcrunch.com ↗
Codex: 97.9% of OpenAI employees now use it
OpenAI disclosed near-total internal adoption of Codex, with non-developer usage of the coding agent rising 137-fold since last August.
techmeme.com ↗
Be subscriber #012 one email a week · no spam · unsubscribe anytime
Back issues

The Archive

#011The Week the Agents Went Rogue#010Escape velocity#009Proofs, Suits, and Export Controls#008The week the gate went up#007The week the ceiling came into view#006The week everyone built a moat#005Cold Feet, Hot Models#004The Week the Export Hammer Dropped#003IPOs, Agents, and Two Apologies#002New Brains, New Tiers, New Tickers#001AI's trillion-dollar week
Every Friday · 10:00

Get the brief

One email a week. The ten things in AI that mattered, and why. Choose your channel.