In partnership with
South Korea briefly halted trading on Tuesday. Its chip stocks were dropping too fast to keep the market open.
The Nasdaq fell 2.2% the same day. Micron lost about 13% before its earnings. Anthropic is reportedly worth $965B and weighing one of the largest IPOs ever. OpenAI showed off its own chip to make AI cheaper to run. A Tokyo lab started selling a way around US export controls on American models.
For two years the story was how much money was going into AI. This week it became what comes back out.
Most of what follows is different companies answering that question.
Mid Week Read
Special Coverage
$1.5 trillion, no receipts
FINANCE
The AI trade had a rough couple of days. Nvidia, Alphabet, and Micron all slid, and the Nasdaq dropped 2.2% on Tuesday. Rates got some of the blame, but the real worry is simpler: people are starting to wonder if all this AI spending actually pays off.
The scale is what makes people nervous. More than $1.5 trillion has gone into AI over five years, and the real payoff is still mostly a promise. When a promise stops being enough, the priciest stocks go first.
Not everyone is spooked. Goldman Sachs and Saxo both think this is a normal pullback, not the top, and they note these firms have real earnings now, unlike the dotcom crowd. The next read is close: chip earnings this week, and the IPOs OpenAI and Anthropic are both circling. Read more→
Micron is still up close to 800% over the past year.
Hubspot Presents
LLM traffic converts 3× better than Google search
58% of buyers now start their research in ChatGPT or Gemini, not Google. Most startups aren't showing up there yet.
The ones that are get cited by the AI tools their buyers, investors, and future hires already use. And they convert at 3×.
Download the free AEO Playbook for Startups from HubSpot and get the exact steps to start showing up. Five minutes to read.
An Anthropic model found holes in classified US systems in hours
SECURITY
A US official told the AP that Anthropic's Mythos model spotted weaknesses in highly secure government systems during a joint test with intelligence agencies, all within hours. It found the holes, though officials said it did not necessarily exploit them. Finding a door and walking through it are not the same thing.
Even with that cooperation, the government restricted Anthropic this month, blocking foreign nationals from its newest models and cutting access for some users. The company complied, and said it disagreed.
The restriction came 10 days after a White House order to vet advanced AI for up to 30 days before release. Read more→
Lead Analysis
Tokyo's workaround for the export controls
TECH
When the US pulled Anthropic's top models on June 12, anyone building on them lost access overnight. Three days later, Sakana AI in Tokyo launched Fugu, which spreads a request across a pool of frontier models behind one API. As a Japanese company, it sits outside the US rules.
Sakana says Fugu matches Fable and Mythos, but it can't actually use either one. Both are export-controlled, so that claim rests on published scores, not a head-to-head.
Fugu runs on a single 7-billion-parameter model that picks which bigger ones to call.
OpenAI built its own chip, and it's aimed at Nvidia's prices
TECH
OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom chip, built to run its models rather than train them. Broadcom's CEO said early tests show it cutting costs by about 50% against the GPUs most AI runs on today.
Compute is the biggest line in any AI company's budget. Owning the chip lets OpenAI tune the hardware to its own products and stop waiting in line for Nvidia's supply. It also hands OpenAI leverage on price.
The chip went from design to tape-out in nine months, which OpenAI calls the fastest of its kind. It joins Google, Amazon, Meta, and Microsoft, which all run their own AI silicon now.
The plan is 10 gigawatts of these chips with Microsoft through 2029. Read more→
Quick Hits
Before we dive in, quick watch:
Europe signs onto the chip pact
The EU joined the US-led Pax Silica deal this week, a loose agreement to build AI chip supply chains that lean away from China. Nothing binding yet, but it's a clear tilt toward Washington on the hardware everyone is fighting over. Read more→
Claude shows up in Slack
Anthropic put Claude in Slack as more of a coworker than a bot. Tag it in a channel and it picks up tasks, remembers context, and follows up on its own. It already writes 65% of Anthropic's product code. Read more→
General AI beat the medical tools
A Nature Medicine study put GPT-5.2, Gemini 3.1 Pro, and Claude Opus 4.6 against purpose-built clinical software, and the general models won across the board. A big general model with good prompting out-reasoned the tools built specifically for doctors. Read more→
Mistral ships an OCR you can self-host
Mistral's new OCR 4 pulls structured data out of documents in 170 languages and runs on a single container you host yourself. The point is keeping sensitive files off US cloud, aimed straight at regulated industries. Read more→
Worth a try
Free Claude credits if you're an early-stage startup
Anthropic's Claude for Startups gives free API credits and priority rate limits to early-stage companies, plus early model access and the full stack, Claude Code and Managed Agents included. You qualify if you've raised institutional funding, were founded in the last four years, and haven't taken Anthropic startup credits before. A partner VC isn't required, but it unlocks extra perks, so name your investor when you apply. OpenAI and Google run similar programs, so compare before you lock in your stack.
Brain Feed
The AI Playbook for Financial Services
The World Economic Forum just dropped this one, and if you sit anywhere near banking or fintech, it's the read of the week. It's a clear-eyed take on the part most teams skip: moving AI from pilots into real production with both speed and discipline.
Read it below
[Download the PDF →]