OpenAI Deepens Bench as Anthropic Fights Washington — AI Daily June 20

It’s the week OpenAI went full “IPO prep mode,” and the rest of the industry didn’t stand still. While one frontier lab scores policy wins, another loses them. While one inference startup doubles its valuation in five months, a hyperscaler admits it can’t make chips fast enough. Let’s get into it.

OpenAI’s Power Play: Shazeer + Ball in 7 Days

OpenAI made two hires this week that tell you exactly where the company’s head is as it edges toward a public market debut.

First: Noam Shazeer, co-lead at Google’s Gemini team and co-inventor of the 2017 “Attention Is All You Need” paper — the Transformer architecture that literally built modern generative AI. Shazeer left Google on Wednesday after the company rehired him in a $2.7 billion deal just two years ago to acquire Character AI’s tech. He’d been at Google since 2000 except for that brief Character AI stint. He’s landing at OpenAI.

Then: Dean Ball, a former Trump White House AI policy official who helped draft America’s AI Action Plan, announced he’ll join OpenAI on July 6 to lead a new “Strategic Futures” team. The mandate covers catastrophic risk, recursive self-improvement, labor market impact, and government-lab relations. He reports directly to Chief Strategy Officer Jason Kwon.

“Almost by necessity, AI labs will have to lead on AI governance decisions.” — Dean Ball, on why internal governance matters more than people realize

Tell you what — if you had “insider policy access” on your IPO risk factor bingo card, this is what it looks like in practice.

Sources: TechCrunch: OpenAI is bringing on some big guns

Anthropic’s Fable 5 Ban: Export Controls Hit a Skeleton Key

While OpenAI builds bench strength, Anthropic is fighting a different battle. Late last week, the Trump administration ordered an export control ban on Anthropic’s Fable 5 and Mythos 5 models, forcing the company to pull them entirely to stay compliant.

The ban targets what OpenRouter calls a “skeleton key” — models that jailbreak other models. The political irony: one frontier lab’s policy connections tighten while another gets squeezed. The numbers so far? Apparently don’t care. Anthropic’s brand appears to be holding.

Source: TechCrunch: The US banned Anthropic’s Fable 5 release

Nobel Winner John Jumper Leaves Google DeepMind for Anthropic

In a separate but equally seismic move, John Jumper — who shared the 2024 Nobel Prize in Chemistry with Demis Hassabis for developing AlphaFold, the open-source protein structure prediction model — announced he’s leaving Google DeepMind for Anthropic.

Jumper had been at DeepMind since 2017. AlphaFold basically solved one of biology’s grand challenges, and the guy who built it is now going to build frontier AI. Anthropic just got a serious research heavyweight in the middle of a government fight.

Source: The Verge

Baseten’s $15B Inference Gold Rush

AI inference startup Baseten is reportedly close to a $1.5 billion funding round at a $13 billion valuation, per the Wall Street Journal. Let that sink in: five months ago, Baseten raised a $300M Series E at a $5 billion valuation. That’s a 160% valuation jump in under six months.

Baseten handles inference by routing requests to the best-for-task model, especially cheaper open-source alternatives. It’s what happens when every company needs AI inference but nobody can get enough GPU capacity.

Fair warning: the round is reportedly “split-priced” — some investors at $13B, others at $11B. A valuation inflation tactic, but still: the inference layer is where the money is flowing.

Source: TechCrunch: Baseten reportedly raising $1.5B

Amazon’s Trainium Gambit: Selling Chips to Everyone Else

AWS is in talks to sell its Trainium AI chips to third-party data centers, a move that would directly challenge Nvidia’s dominance. Peter DeSantis, Amazon’s AI chief, told Bloomberg the company is exploring selling racks externally.

CEO Andy Jassy put a number on it in April: if Trainium were a standalone business, its annual run rate would be ~$50 billion. Currently, Trainium capacity sells out instantly — Jassy said even Trainium4 (not available for over a year) is already pre-sold.

The catch? TSMC capacity. Nvidia just supplanted Apple as TSMC’s largest customer, so Amazon will need to elbow its way into the foundry queue. This is the chip equivalent of “easy to say, hard to build.”

Source: TechCrunch: Amazon hopes to challenge Nvidia

The Big Picture

This week’s pattern: poaching, policy positioning, and infrastructure bets. OpenAI is assembling an all-star team (technical + regulatory) for the IPO runway. Anthropic is weathering government headwinds while landing a Nobel-caliber researcher. The inference layer is attracting massive capital. And the chip wars are about to get a new competitor with very deep pockets.

The frontier labs aren’t just building models anymore — they’re building moats. Talent moats, policy moats, infrastructure moats. The spaghetti is getting thicker.

Affiliate links disclosure: Some links in this post may be affiliate links. We may earn a small commission if you purchase through them, at no extra cost to you. This doesn’t affect our editorial independence.