Week 24 Roundup — The AI Industry Is Repricing Everything

The Big Picture

This week had one through-line: the AI industry is repricing everything at once — compute, tokens, talent, and trust. SpaceX went public at a valuation that only makes sense if orbital datacenters become real. Anthropic launched its most capable model and immediately alienated its enterprise customers. OpenAI is considering a price war. Google is renting Nvidia GPUs from a rocket company. Every major actor is making a bet that the current pricing, infrastructure, and trust dynamics are about to change — and making those bets in opposite directions.

The most important signal of the week wasn’t the SpaceX IPO or the Fable launch. It was the Codex usage spike after Fable’s backlash. That’s the market telling you that enterprise customers are not captive, they can and will move, and the companies that assume lock-in are going to learn the same lesson Coinbase learned when AWS went down and there was no failover.

Best Of This Week

1. Anthropic’s Fable Backlash — The AI DR Exercise — June 11 & 12

The most important story of the week, and the one with the longest tail. Fable 5 launched with two policies that enterprise customers rejected immediately: mandatory 30-day data retention with no opt-out, and silent model degradation for anyone working on LLM-adjacent infrastructure. Microsoft blocked it internally. AWS confirmed it breaks their security boundary on Bedrock. Codex saw an immediate traffic spike.

Anthropic walked back the silent nerfing after 24 hours of backlash — they’ll now make degradation visible — but held firm on data retention. The principle stands: Anthropic is establishing the infrastructure and contractual precedent for retaining enterprise traffic, today framed as “safety only, no training.” That’s a foot in the door.

The frame I keep coming back to: enterprises need AI disaster recovery plans now. Can your team failover from Anthropic to OpenAI in 24 hours? If the answer is no, you have a single point of failure in your stack that’s more fragile than most infrastructure teams would accept for any other dependency. Model routing and provider abstraction aren’t a nice-to-have — they’re the same discipline as cross-zone failover.

2. Siri Finally Works — And the Index Is Why — June 9

The most underreported WWDC story was buried in the keynote: Apple rebuilt their search index. Siri’s failure for two years wasn’t purely about bad models — it was that searching your own Messages and Mail was broken at the infrastructure level. You can’t build personal AI on a broken search layer. That’s fixed now, which is why the demos actually worked for the first time.

The bigger strategic picture: Apple is running cloud inference on Nvidia H100s inside Google Cloud, not Apple Silicon. The Google/Gemini deal for $1B/year is actually a Google Cloud customer acquisition. Google is likely using SpaceX’s Nvidia GPUs for internal workloads while freeing up capacity for Apple — and simultaneously steering enterprise customers onto TPUs for lock-in. Everyone in this story is playing a longer game than the headlines suggest.

3. The SpaceX IPO and the Orbital Compute Thesis — June 11 & 12

100x revenue is an insane valuation by any traditional metric. The bull case isn’t the current financials — it’s the vertical integration story. Reusable rockets make orbital datacenters viable. Orbital datacenters solve the two walls the industry keeps hitting: power and cooling. In space you radiate heat into vacuum; no cooling infrastructure required. That’s a structural advantage nobody else can replicate without also building a launch business.

The unsolved problem nobody is talking about: maintenance. Terrestrial datacenters have humans. In orbit, if something breaks it becomes space junk. The economics only work if failure rates stay low enough to depreciate and replace faster than hardware fails. That’s the engineering problem that determines whether this thesis holds or becomes the most expensive science experiment in history.

One Thing I’m Watching

Smart model routing is about to become the enterprise AI cost management category. Token prices between frontier and mid-tier models can differ by 10-20x. The Pragmatic Engineer this week catalogued a dozen vendors — Factory Router, Not Diamond, OpenRouter, Weave — that are already claiming 20-30% cost savings by routing queries to the right model automatically. Cloudflare AI Gateway has the infrastructure to do this today but wasn’t mentioned in the list. If the product team positions it as a routing and cost optimization layer, not just an observability tool, there’s a real market there. Worth watching whether that positioning shift happens before the routing space consolidates around a winner.