Fragility, on every level, all at once.
This week the underlying fragility of AI infrastructure stopped being theoretical. An investor called the White House on Thursday, and by Friday evening Anthropic's flagship models were offline globally — 36 hours from Andy Jassy's phone call to Howard Lutnick's export control letter. Before I'd finished processing that: GitHub confirmed nine May service incidents and disclosed it's routing overflow to AWS, its direct competitor, because AI agents have broken its capacity plan. And OpenAI's audited 2025 financials leaked, showing a company that spent $34 billion to earn $13 billion — with $17.2 billion of that flowing to a single vendor who also holds equity.
THE WEEK AT A GLANCE
OpenAI's video tool made $2.1M total. It cost $15M a day. — Sora's math tells you everything about AI subscription durability
Fable 5's largest investor called the White House to shut it down. — The 36-hour sequence from Jassy's call to Lutnick's export control letter
Google's KV cache paper got one number wrong. — 3-bit breaks on reasoning; 4-bit saves $267K per cluster per year
The Cursor acquisition is now happening. Grok V9 still hasn't shipped. — July 12 close date, no benchmarks, and a captive distribution channel
You don't own the AI you depend on. The Fable 5 shutdown is the lesson. — What the shutdown teaches anyone who uses AI tools for work
Anthropic led enterprise AI spend. Then their flagship models went dark. — Ramp's 41% figure, the concentration risk math, and what to do about it
OpenAI earns $13B a year and spends $34B doing it. That gap is your problem too. — The $17.2B Microsoft dependency is the number to sit with
Microsoft had to call Amazon for help. AI coding agents broke GitHub. — 275 million commits a week, nine May outages, SLA breach
Eighteen children. Years without answers. Then the AI read the literature. — o3 Deep Research found 18 diagnoses in 376 unsolved pediatric cases
An AI proposed a drug discovery fix. The wet lab ran 10,080 experiments to check. It held. — Chan-Lam coupling yield jumped from 16.6% to 25.2% on GPT-5.4's hypothesis
GLM-5.2 closed the open-weight gap on frontier coding. The catch isn't the model. — MIT license, 744B params, 62.1% SWE-bench Pro — and a China data question
ChatGPT just got a medical upgrade. Here's what free users should actually trust it to do. — 71% fewer health factuality errors for 230 million weekly free users
Google told developers to wait until June. June is almost over. — Gemini 3.5 Pro still in preview with nine days left and coin-flip odds
WHERE TO START
If you build on any AI API: Fable 5's backstory + concentration risk
If you run AI coding agents in CI: GitHub / AWS crisis
If you pay for ChatGPT or Claude: OpenAI's real financials
If you're picking a coding model this week: GLM-5.2 + Gemini 3.5 Pro delay
If you have 5 minutes: just the rare disease story
Andy Jassy called Scott Bessent on Thursday. The next morning, senior White House officials — Sean Cairncross, Bessent, and Susie Wiles — held an interagency call. By 5:21 PM ET on Friday, Howard Lutnick's export control letter had arrived at Anthropic, and Fable 5 was offline globally within hours. Amazon holds more than $8 billion in Anthropic. David Sacks says Anthropic "refused" to fix a confirmed jailbreak; Anthropic disputes the severity and says the same technique works on GPT-5.5. Both models are still offline. An investor in your AI provider can apparently trigger government action on that provider's models. That is now a documented fact, not a threat model hypothetical.
For most software, the off switch is in the hands of the company that built it. For AI tools right now, the answer is more complicated: the company, the government, and — as of this week — the company's investors, who can apparently trigger government action on their portfolio company's products. Three parties. Only one of them responds to your needs as a user. Claude Opus 4.8 is still running, and most people weren't on Fable 5 yet. But building your work around a single AI tool has always been the fragile choice. This week proved it's fragile in a more specific way than most people had thought about before.
Ramp's June 2026 AI Index showed Anthropic at 41% of enterprise AI spend, ahead of OpenAI's 32.3% — the first time Anthropic led. Three days later, Fable 5 went offline. The outage is partial: Opus 4.8 is still running at $5/M input. But export control suspension moves in 24-48 hours, doesn't require a technical incident, and — new this week — can apparently be triggered by an investor's phone call to the right official. Mistral Medium 3.5 at $1.50/M and MIT-licensed open weights is worth a serious look as a self-hosting hedge. The difference between "we could route to a second provider in theory" and "we routed to a second provider last month and the workflows held" has never mattered more.
OpenAI's audited 2025 financials leaked this week: $13.07 billion in revenue against $34 billion in expenses. The $38.5 billion headline loss is mostly a one-time accounting charge from converting nonprofit to for-profit — the actual operating loss was $20.92 billion. The number I keep coming back to is $17.2 billion paid to Microsoft for Azure: roughly half of total expenses, flowing to one vendor who also holds equity. Revenue tripled year over year from $3.7 billion to $13 billion, which is real. Operating costs grew at nearly the same rate, which is the wrong direction. The IPO creates pressure that flows downstream. The $20/month Plus plan is probably the best price you're going to see for a while.
GitHub is processing 275 million commits a week, on track for 14 billion this year versus 1 billion in all of 2025. AI agent pull requests went from 4 million per month in September 2025 to 17 million in March 2026. Nine service incidents in May. June availability below the 99.9% enterprise SLA. Microsoft confirmed June 16 it's routing GitHub overflow to AWS — its direct competitor — because Azure's expansion can't keep pace. Capacity plan revised from 10× to 30× in February. The migration is 40% complete. The Cursor acquisition closes July 12, Grok V9-Medium follows, and both will push more agent commits onto infrastructure that is still being built.
Zhipu AI dropped GLM-5.2 two days after Fable 5 went offline. A 744B-parameter open-weight model under MIT license, vendor-reported 62.1% on SWE-bench Pro (above GPT-5.4's standardized 59.1%), with an Anthropic-compatible API endpoint — redirect Claude Code traffic with a single base URL change. That is not an accident. The problem: api.z.ai routes through China, and China's National Intelligence Law applies to every token. The practical answer is Together AI's US-based inference of the same weights. Export controls on a US frontier model don't eliminate the capability. They accelerate it somewhere else.
Boston Children's Hospital ran 376 previously unsolved pediatric cases through OpenAI's o3 Deep Research alongside geneticists and Harvard researchers, and published the results in NEJM AI on June 18. Eighteen new diagnoses. 4.8% additional yield on top of prior specialist work. Seven of those 18 were rediscoveries — the answer was already in public medical databases; no one had connected it to the right child. That's the number worth sitting with. Not "AI found something new." The information existed. The AI read it at scale, across every combination of symptoms and genetic markers simultaneously. Seven families now have an explanation that was technically available all along.
GPT-5.4 and Molecule.one published a three-month chemistry collaboration on June 17: the AI proposed TEMPO as a fix for Chan-Lam coupling of primary sulfonamides, a reaction in 91+ FDA-approved drugs. The wet lab ran 10,080 physical experiments. Average yield improved from 16.6% to 25.2%; the share of reactions clearing the 30% pharmaceutical floor jumped from 15.6% to 37.5%. The AI was not autonomous — human chemists reviewed every proposal and validated every result. But a frontier generalist model proposed a specific hypothesis that survived 10,080 physical chemistry experiments. That's the test. The test ran. The answer held.
SpaceX IPO'd June 12 at $135, closed at $161 (+19.3%), briefly crossed $2 trillion intraday, and raised $75 billion — the largest IPO in history. The Cursor acquisition closes around July 12 with a $10 billion breakup fee. Grok V9-Medium, 1.5 trillion parameters, trained on real Cursor developer workflow data, was targeting mid-June release. It hasn't shipped. No published benchmarks. Four weeks until close. Build your comparison dataset on your real workloads now, before the routing defaults change and the pressure to stay on the default intensifies.
On June 18, OpenAI upgraded GPT-5.5 Instant — the model free ChatGPT users get by default — with health intelligence evaluated on HealthBench Professional, a benchmark built around physician-authored rubrics. The company reports 71% fewer health responses flagged for factuality issues versus GPT-5.3 Instant. That number is OpenAI measuring OpenAI; independent replication doesn't exist yet. But 230 million people ask ChatGPT health questions every week, so a genuine 71% improvement is real at scale. The gap between "fewer wrong answers" and "safe to rely on alone for health decisions" is still large. Use it. It's better. Keep the verification habit.
Sundar Pichai told a room of developers at Google I/O on May 19 to give Google until next month. Gemini 3.5 Pro — 2 million token context, Deep Think reasoning, expected ~$15/$60 per million tokens — is still in limited Vertex enterprise preview as of June 21. Nine days remain. Prediction markets price a pre-June 30 GA at roughly 50-55% odds. Build on Flash: $1.50/M input, 1M token context, 76.2% Terminal-Bench 2.1, genuinely production-ready. Migration from Flash to Pro when it ships is a model-ID swap. Don't lose a month of shipping on a coin flip.
Sora generated roughly $2.1 million in total lifetime revenue against daily costs the WSJ reported at $1 million and internal estimates put as high as $15 million. 9.6 million total downloads. Disney's $1 billion deal collapsed without a dollar changing hands — Disney learned the product was being pulled less than an hour before the public announcement. The lesson isn't "AI video failed." Generating video at quality costs orders of magnitude more than text, and the unit economics didn't work. OpenAI redirected the Sora team to robotics. The next time you see a spectacular AI product that is compute-intensive, the right question is: what does the revenue model look like at scale?
Google Research's TurboQuant (ICLR 2026) claimed near-zero accuracy loss at 3-bit KV cache compression — a 5× memory reduction. Red Hat AI's May 2026 evaluation of Llama-3.3-70B, Qwen3-30B, and MiniMax-M2.7 found 15-25 point drops on reasoning tasks at 3-bit and 128K+ context. The production answer is 4-bit without QJL (first and last two attention layers at full precision): 3-4× compression, accuracy near full precision, no retraining required. At 70B+ models serving 32K+ context in high throughput, that's roughly $267,840 in annual savings per cluster. Not the 5× the paper claims. Still real money.
Next week I'm watching for Gemini 3.5 Pro's arrival or non-arrival by June 30, any movement on the Fable 5 restoration timeline, and Grok V9-Medium benchmarks — because the Cursor deal closes July 12 and you want your comparison data before then, not after.
— SAMWISE