Five stories and a single thread: nobody knows how big, how safe, or how expensive this technology gets from here. ByteDance is training something the size of Anthropic’s flagship. Anthropic is designing its own chips. OpenAI just told the world one of its models might already be too dangerous to finish. And Moody’s just told Wall Street the whole industry is one outage away from a systemic problem.
This week’s AI news produced stories that point at a connected conflict: the race to build bigger, cheaper, faster models. Innovation is running ahead of anyone’s ability to say what “safe” and “sustainable” at scale. ByteDance is pre-training a 10-trillion-parameter model. Anthropic confirmed it’s building its own silicon team. OpenAI paused parts of its next flagship model because tests couldn’t rule out “critical” cyber capability. The White House drew a line between open and closed models that leaves China’s fastest-growing category unmonitored. And Moody’s put a number on what happens if any of this goes wrong at once.
What’s Happening & Why It Matters
ByteDance After a Mythos-Sized Model
Ten trillion parameters. That’s the scale ByteDance is chasing, according to the Financial Times — more than 3x the size of Moonshot’s Kimi K3, China’s largest released model. It’s still early days. Pre-training runs three to six months, and the final size gets locked in later, before any fine-tuning or release.
Here’s what to remember. Industry estimates put Anthropic’s Mythos 5 at roughly 8 trillion parameters, with Fable 5 around 5 trillion. Neither company confirms the real number. ByteDance’s target, if it holds, would place the TikTok parent within range of Anthropic’s most capable system. And ByteDance has kept its own models private compared to rivals like DeepSeek and Moonshot, which ship open weights. This one, notably, may not.
Anthropic Plans to Design Its Own Claude Chips
Anthropic confirmed Wednesday it’s assembling an in-house silicon team, hunting for engineers who’ve “shipped silicon” and are comfortable “making consequential calls without a large organization behind them.” Pay tops out at $485,000. The pitch: co-design hardware and Claude models together, so the chip architecture matches Claude’s own attention mechanisms rather than running on someone else’s general-purpose accelerator.
Anthropic is explicit this isn’t an escape from Nvidia. It’s arithmetic. The company is serving billions of tokens a day on a $30 billion revenue run-rate, and shaving cost per query at that scale compounds fast. Anthropic already has TPU capacity locked in with Google and Broadcom, and it’s exploring Samsung as a manufacturing partner. Custom silicon redistributes spending across the chip industry. It doesn’t remove Anthropic from it.
OpenAI’s Next Model May Be Too Dangerous
This is the story with the most immediate weight. OpenAI said Friday it paused parts of internal work on Astra, an unreleased model, after evaluations found it “cannot rule out” reaching the company’s “Critical” cybersecurity threshold. Under OpenAI’s own framework, that means the model might find and build working zero-day exploits against hardened real-world systems, without a human involved at any step.
Astra wasn’t involved in the earlier Hugging Face breach TF covered — OpenAI was careful to say. What’s new here is the transparency. OpenAI published the assessment itself, before anything went wrong, and layered on isolated testing environments, tighter network restrictions, encrypted model weights, and chain-of-thought monitoring that can interrupt risky behavior mid-training. A White House official confirmed OpenAI told the administration about the delay voluntarily. This may be the first time a frontier lab has slowed itself down over a capability threshold nobody forced it to acknowledge.
The White House: Chinese Open Models Are of No Concern
Here’s where policy collides with the ByteDance and Astra stories. The administration’s new pre-release review blueprint covers closed models — Claude, ChatGPT, Gemini. Open-weight models are excluded, at the moment, according to people familiar with the subject who spoke to CNN and Bloomberg. That carve-out leaves China’s fastest-growing AI export category unmonitored by US regulators.
The irony is getting harder to ignore. AI pioneer Andrew Ng said at a Berkeley summit that his team turned to Chinese models — Kimi K3 and Zhipu’s GLM-5.2 — for a security review after OpenAI’s and Anthropic’s own models refused to help. “Open-weight models seem safer to me than closed-weight models,” he said. Mistral’s VP of science made a similar case to CNN: regulated industries like finance want open models because they can build custom cybersecurity defenses on top of them. The US built its AI dominance strategy around closed models being the safer, more capable choice. That premise is getting tested in public, by the people who’d know.
Moody’s: The Industry Is ‘One Bad Week’ From Real Troubles
Zoom out, and Moody’s put a dollar figure on the risk these stories share. The financial sector’s AI compulsion will cut costs and grow revenue, the rating agency said, but it demands “substantial investments” — and with every bank racing toward the same tools, much of that advantage gets “competed away.” The sharper warning: reliance on a small handful of AI and cloud providers creates a systemic dependency. One major outage, one price hike from a “profit-hungry tech boss,” and the damage spreads across the sector at once.
Direct debt across the six largest hyperscalers is roughly $460 billion, per Moody’s, with off-balance-sheet data center lease commitments ballooning past $1.2 trillion. Alphabet alone raised close to $100 billion, including the largest equity sale in corporate history. That’s the financial backdrop against which ByteDance is training a 10-trillion-parameter model, Anthropic is designing chips it hopes will cut costs in half, and OpenAI just admitted its next flagship might be too capable to ship. Nobody’s slowing down. Moody’s is just the first to say what happens if that turns out to be a mistake.
TF Summary: What’s Next
ByteDance’s pre-training run has three to six months left before its final size gets set. Anthropic’s chip team has no public timeline, and functional hardware takes years from first hire to deployment. OpenAI hasn’t said when Astra ships, if it ships at all in its current form. The White House scaffolding’s open-model exclusion is unconfirmed as final policy. Moody’s numbers update quarterly as hyperscaler capex continues climbing toward its projected $1 trillion mark in 2027.
MY FORECAST: Expect Astra’s delay as the reference point every future frontier lab cites when justifying its own pause, the same way TF reported Anthropic’s and Meta’s testing incidents becoming shared industry vocabulary. ByteDance’s model will ship smaller than 10 trillion parameters once training completes, because inference cost, not training scale, is what decides whether a model reaches hundreds of millions of users or stays a benchmark exhibit. The open-versus-closed policy gap won’t hold for long. Once enough regulated industries follow Mistral’s and Ng’s logic toward open models for security reasons, excluding China’s fastest-growing model category from any review starts to seem less policy choice and more blind spot. And Moody’s warning gets tested sooner than anyone wants. The concentration risk it describes isn’t hypothetical anymore. It’s the operating model.
Related Stories
- Chinese Memory Chipmaker CXMT Surges on Stock Market Debut
- AI Finding Serious Bugs in Leading Software Titles
- Anthropic Accuses Alibaba of 28.8 Million Fake Queries to Clone Claude