In partnership with
In Todayโs Issue:
๐ How an Amazon warning helped take Anthropic's top models offline
๐ง Moonshot's Kimi K2.7 Code, an open trillion-parameter coder
๐ NVIDIA and Apple bring confidential computing to cloud AI
๐๏ธ Rio's city government ships its own open 397B model
๐ What Epoch's corrected FrontierMath scores really reveal
๐ธ SpaceX's record run at the biggest IPO in history
โจ And more AI goodnessโฆ
โก The Signal
A model can be the most powerful product in tech on Monday and switched off by the government by Friday.
That is the lesson still echoing from Anthropic's lost weekend. New reporting this week says it was a warning from Amazon, not a regulator, that pushed the White House to disable Fable 5 and Mythos 5, and three days later the models are still dark. The deeper pattern is a split screen. As Washington starts treating the best closed models as strategic assets it can switch off by letter, the open-weight world is sprinting the other way: China's Moonshot just shipped a trillion-parameter coder anyone can download, Rio's city hall released its own open model, and NVIDIA is hardening the privacy of cloud AI. One frontier is becoming locked and sovereign; the other, open and borderless.
All the best,
Kim Isenberg
๐ง China's Moonshot Ships an Open Trillion-Parameter Coder
Moonshot AI has released Kimi K2.7 Code, a one-trillion-parameter open-weight model built for hands-off software work. It uses a mixture-of-experts design that activates only 32 billion of its 1T parameters per request, runs a 256,000-token context (the amount of text it can weigh at once), and ships under a modified MIT license that lets anyone download and self-host it. Moonshot says it lifts its Kimi Code Bench v2 score to 62.0 from the prior 50.9 while cutting "thinking" tokens by about 30%, a push to make capable coding agents cheaper to run.
๐ tl;dr: As the US locks down its closed frontier models, China's open-weight labs keep shipping the opposite: strong code models anyone can run.
๐ NVIDIA and Apple Bring Confidential Computing to Cloud AI
NVIDIA says its Blackwell GPUs can now run Apple's Private Cloud Compute, the privacy system behind Apple Intelligence, on Google Cloud. "Confidential computing" seals each AI request inside a hardware-locked "trusted execution environment" and cryptographically proves the servers have not been tampered with, so that, in Apple's words, "no one, not even the system's builders, can look at their data." Announced at Apple's WWDC, it lets heavy server-side models run with the same privacy promise Apple makes on your phone.
๐ tl;dr: The privacy fight is moving into the data center, where the goal is AI that is powerful in the cloud yet unreadable to the people who run it.
๐๏ธ Rio's City Hall Releases Its Own Open AI Model
The city government of Rio de Janeiro has published Rio-3.5-Open-397B (yes, you read that correctly), a 403-billion-parameter open model that works in Portuguese and English. Released on Hugging Face under a permissive MIT license, it was built by merging two existing open models and then fine-tuning, and it has already passed 112,000 downloads. The team openly flagged that an earlier upload was the wrong, unfinished version and said it is re-uploading the final model, a candid reminder that public-sector AI is still finding its feet.
๐ tl;dr: When a city hall can ship a 400-billion-parameter model, frontier-scale AI is no longer just a big-lab privilege.
Stop single-sourcing your AI stack.
Why it helps: Friday showed that a frontier model you depend on can vanish overnight, not from an outage but from a government letter. Teams with a second model already wired in kept working; everyone else just waited.
Try this: Paste this into your AI: "List every workflow where I rely on a single AI model. For each, name one credible alternative, including an open-weight model I could self-host, and the exact steps to switch over within a day if my main model goes dark for a week."
๐ฌ Watch This
In this video, financial educator Ben Felix breaks down everything about the recent SpaceX IPO, exploring how recent rule changes by major indices will impact whether the stock is automatically added to your passive index funds.ย He also dives into the historical data of tech IPOs to answer whether everyday investors should buy in directly, highlighting why long-term returns often fall short of their initial first-day hype. It is a clear-eyed antidote to IPO fever, well timed as Anthropic and other AI giants edge toward public listings of their own.
โ Jensen Huang, NVIDIA CEO, at CES 2026, where he noted that roughly 80% of startups now build on open models.
The trigger for Washington's off switch may be quieter than the jailbreak story suggests: China.
According to The Verge and Semafor, the export controls were imposed partly over suspicions that a China-linked group had gained access to Mythos, Anthropic's unrestricted, defenders-only cyber model, and that this, more than any single demo, is what set off Friday's emergency shutdown of Fable 5. The reporting is carefully hedged. Semafor stresses it is "unclear how the White House learned of the issue, which organization accessed the model, and how it gained access to Mythos." Anthropic says the White House never raised Chinese access in their talks, and AI czar David Sacks has pointed only at the Fable 5 jailbreak, not China. For now it is a serious but unconfirmed thread, not a settled fact.
How an Amazon Warning Took Anthropic's Models Offline
The Takeaway
๐ On Friday, June 12, a US export-control letter forced Anthropic to disable its two most powerful models, Fable 5 and Mythos 5, for every foreign national, so to comply it switched them off worldwide. Three days on, they are still dark.
๐ New reporting says the trigger was Amazon: CEO Andy Jassy warned the White House that researchers had bypassed Fable 5's guardrails to pull restricted information about cyberattacks.
๐ It set off a frantic 24 hours, with Treasury Secretary Scott Bessent telling CEO Dario Amodei he was making a "bad decision" before Commerce Secretary Howard Lutnick sent the order.
๐ Anthropic calls the flaw a narrow jailbreak that rival models share; AI czar David Sacks says it is serious and that "the ball is in Anthropic's court." The models stay offline as talks continue.
First, a quick recap, because the models vanished on a Friday night and the story has moved fast since. On Monday, June 9, Anthropic launched Fable 5 and Mythos 5, calling Fable the most powerful AI model ever released to the public (Mythos is its unrestricted twin, shared only with vetted cyber-defenders). Four days later they were gone. Citing national-security law, the Commerce Department ordered Anthropic to bar all foreign nationals from the two models, even its own non-citizen staff inside the US. Because the company cannot check every user's passport in real time, the only way to comply was to switch both models off for everyone, which it did over the weekend; Amazon Web Services pulled hosted access too.
Anthropic CEO Dario Amodei, who fielded three tense White House calls before the shutdown (Bloomberg via TechCrunch)
What is new this week is how it happened. According to Politico and Fortune, the alarm came not from a regulator but from a rival: Amazon CEO Andy Jassy, whose researchers found they could prompt the Mythos-class model into handing over restricted details about cyberattacks. He flagged it to senior officials on Thursday. By Friday morning the White House was on a call, Treasury Secretary Scott Bessent dialing in from Houston, and when Dario Amodei finally joined he faced three tense calls with about half a dozen officials. Bessent told him bluntly he was making a "bad decision." As one senior official put it, "Export controls were a last resort after begging them for hours to work with us."
The White House moved within 24 hours to restrict Anthropic's models (Getty Images via Fortune)
The two sides still disagree on almost everything, and the models are still down. Anthropic argues the jailbreak is "narrow and non-universal," the kind of thing rival models like OpenAI's GPT-5.5 will also do, and that recalling a product used by hundreds of millions over one demo sets an impossible standard. Sacks counters that a "highly credible, trusted partner" found the flaw, that Anthropic declined to fix it, and that "the ball is in Anthropic's court." Over the weekend the company sent senior technical staff to Washington to negotiate. The stakes are high: Anthropic filed confidentially for an IPO on June 1 near a $965 billion valuation, and its most valuable products, it turns out, can be switched off by letter.
Why it matters: For the first time, the deciding vote over who may use a frontier AI model sat with the government, not the company that built it, and we now know the lever can be pulled within hours of a tip from a competitor. Every lab planning to ship something more capable than the last just learned that national-security power, not the market, may decide whether it stays switched on.
Sources:
๐ https://www.axios.com/2026/06/12/anthropic-trump-mythos-fable-national-security
๐ https://www.wsj.com/tech/ai/anthropic-halts-access-to-top-ai-models-after-u-s-ban-on-foreign-use-a4bca2cc
๐ https://www.politico.com/news/2026/06/13/inside-the-whirlwind-24-hours-that-led-the-white-house-to-slap-export-controls-on-anthropic-00961519?utm_medium=twitter&utm_source=dlvr.it
The GTM bets that shouldn't have worked, and did
One grew revenue 50x after half his team quit over the strategy. One brought in 50K signups in a single day with no paid budget. One generated 100M+ views from a stunt that took 50 hours to conceive. One asked every prospect to demo the product themselves instead of demoing it for them.
None of them followed the safe playbook. They treated GTM like an experiment, moved before they had proof, and made bets most founders would never get approved.
HubSpot for Startups documented all 6 stories in the free Bold Bets Playbook. The risks they took, why it was risky, and what it returned.
The chart: From Epoch AI, comparing how the GPT-5 series (GPT-5, 5.2, 5.4, 5.5) scores on FrontierMath, a benchmark of research-level math problems, before and after Epoch corrected errors in the dataset. On every model the corrected (blue) scores beat the original (teal). On the easier Tiers 1 to 3, GPT-5.5 climbs from about 50% to 80%; on the hardest Tier 4 it jumps from roughly 38% to 71%, and even GPT-5 doubles from about 10% to 20%.
The lesson: Some of what looked like the models failing was really the test being wrong. When Epoch fixed flawed problems and answers, measured ability rose sharply, most of all on the hardest questions. A benchmark is a measuring instrument, and a miscalibrated one quietly understates what today's models can already do.
The caveat: This covers only the GPT-5-series (non-Pro) models that were state of the art at release, run at maximum reasoning effort on Epoch's own scaffold, and "corrected" means Epoch's own revisions, not a neutral referee's. The scores still are not 100%, and there is no guarantee other model families would shift the same way.
๐ธ SpaceX's IPO Is Shaping Up to Be the Biggest Ever
โก Bottom line: SpaceX is reportedly heading for a 2026 public listing at a valuation around $1.5 trillion, on track to be the largest IPO in history.
๐ก Why it matters: It is really a bet on Starlink, the satellite network that has become core connective tissue for AI systems, drones, and defense.
๐ What it means: Retail investors may get an unusually large slice of the offering, and with it an unusually large chance to overpay for the hype.
The year's biggest finance story may not be an AI lab at all; it is the rocket company that wants to go public. Reports say SpaceX is preparing a 2026 IPO that would value the whole company, Starlink included, at roughly $1.5 trillion or more and raise on the order of $25 billion to $30 billion. That would edge past Saudi Aramco's 2019 debut to become the largest IPO in history. A recent secondary sale already valued SpaceX north of $800 billion, with employees able to cash out at about $420 a share.
SpaceX's Starship on the pad during a full-pressure test (SpaceX via TechCrunch)
The AI angle is Starlink. SpaceX's value rests less on rockets than on the world's largest satellite network, which is fast becoming the connective tissue for AI: always-on bandwidth for autonomous systems, drones, sensors, and far-flung data centers. Tie that to Elon Musk's other bet, the xAI lab now woven into X, and a SpaceX share starts to look like a proxy for a whole Musk AI-and-space empire. The offering is unusual in another way too: reports suggest SpaceX may hand as much as 30% of it to retail investors, far above the typical 5 to 10%.
Elon Musk, whose SpaceX and xAI bets are increasingly intertwined (The Sun)
The caution is the same one this week's Anthropic saga taught: scale invites the state. Senator Elizabeth Warren has already urged the SEC to halt the SpaceX listing on governance and national-security grounds, a reminder that a company this strategic does not float entirely freely. For ordinary investors it is a rare chance to buy into a trillion-dollar story at the open, but today's Watch This lays out why first-day hype so often outruns the long-run return.
No theory. No slides. Just pipeline.
Most founders know their product. Few know how to get it in front of the right people. In this hands-on session, Clay + HubSpot for Startups walk you through ICP definition, prospect list enrichment, and AI-personalized outreach. You launch your first sequence before the session ends. June 18. 11am ET / 4pm GMT.