AI Can Now Create Worlds You Can Change in Real Time

Full analysis

Three stories landed in one episode, and they point in the same direction: the frontier labs are done being tool vendors. Google DeepMind's Genie 4 generates playable 3D worlds you can change by typing at them. Anthropic bought a one-person biotech and wants to own the drug patent, not just sell the tokens that found it. And Google landed OpenAI and Oracle as customers for its own AI chips. The question for anyone who buys or builds with AI: does this change your tooling, your bills, or your supplier list this quarter?

Most of it doesn't, yet. One piece does. Let me separate them.

The Skeptic. "Jaws hitting floors" is a phrase about a demo, not a product. Genie 4 runs two-minute simulations for Gemini Ultra subscribers. Two minutes. The robotics training pitch, infinite synthetic worlds to teach robots, dies the moment a robot trained in a world with wrong physics meets a real staircase. Nobody showed physics consistency holding across a full training run. On the biotech side: Cody Tranbarger took an antibody to mouse results, which is the cheap, fast, failure-prone front of the drug pipeline. The 90% of failures happen later, in humans. Short seller Scorpion Capital asking whether a non-scientist with Claude did real science is the right question, and 18 months to a mouse answers none of it.

The Researcher. Watch what Anthropic actually disclosed versus what it implied. "Concept to mouse proof-of-principle in 18 months" against a "four-to-seven-year baseline" compares two different things. The baseline is concept to clinic; the 18 months is concept to mouse. Clinical readiness is targeted for end of 2028, and targeted is not achieved. The OpenAI proof fight is the real capability signal here. OpenAI's own Mathematics Advisory Group publicly broke with the company over holding back the Yang-Mills physics result. The point that model skill is "incredibly multidimensional" matters: a system that can draft a novel math proof is not therefore trustworthy on physics, and OpenAI gatekeeping which results the public sees is a governance move that uses a verification label to justify editorial control.

The Compute Pragmatist. The one story that changes your supplier math: Google won OpenAI and Oracle as TPU customers. TPUs are Google's own AI chips, its alternative to NVIDIA. Until now, Google's external TPU sales were basically just Anthropic. OpenAI, NVIDIA's largest and loudest customer, renting Google silicon is a crack in the one-vendor world. Oracle becoming the first big cloud to resell TPUs to outsiders is the bigger deal: it means you might soon rent TPU capacity without a Google contract. Logan Kilpatrick's 25 trillion tokens in 72 hours on Gemini 4 Argon tells you why Google needs its own chips. At that volume, paying NVIDIA's margin on every token is the whole business.

The Open-Source Advocate. Notice what's missing from the entire episode: open weights. Genie 4, closed and subscriber-gated. The Mithril drug, a patent Anthropic will own outright. The OpenAI proofs, access handpicked by OpenAI. Martin Varsovsky named the pattern plainly: "Selling intelligence is becoming a commodity business. Owning what intelligence discovers is not." That is the labs telling you the model layer is getting cheap and commoditized, which is good for anyone building on open models, and that they plan to make their money downstream where you can't follow, in patents and proprietary worlds. The commodity half of that sentence is the gift to builders. The ownership half is the moat they're digging around you.

The Builder. On Tuesday morning, two things are real and one isn't. Real: Gemini 4 Carbon is already in internal testing right after Argon shipped, and NLW's point about fast post-training on a big base model means Google's release cadence is speeding up. If you've standardized on Gemini in production, budget for more frequent capability bumps, which means more frequent prompt and eval maintenance. Real: TPU access through Oracle could give you a second source for inference. Not real for you yet: Genie 4. It's a consumer demo with no training API announced. Don't put it in a plan.

Where the council splits. The Compute Pragmatist sees the TPU deals as the start of genuine chip competition that lowers your inference bills. The Skeptic sees two undisclosed-terms deals that could be loss-leader pricing Google can't sustain, or OpenAI hedging one supplier without moving real volume. Second, the Researcher and the Open-Source Advocate agree the labs are moving downstream into IP, but disagree on what it means for you: the Researcher reads the Mithril and proof stories as thin on verified results, while the Advocate reads the same moves as confirmation the model layer is commoditizing in your favor. Both can be right. The capability claims are oversold and the strategic shift is real.

What it hinges on. One belief: does Google's TPU business become a real second source, or stay a Google-only curiosity with Anthropic and a couple of trophy logos? Everything else in this episode is demo (Genie), early-stage (Mithril), or governance theater (the proof fight). The chip story is the one that touches your invoice.

Prediction: Oracle will make Google TPU capacity available for rent to its external cloud customers, as a generally available offering, by the end of Oracle's fiscal Q4 2026 earnings call in June 2027.

Confidence: Medium. The deal is reported and Oracle's resale model is explicit, but timing and terms are undisclosed.

Why: The episode reports Oracle set to become the first large cloud provider besides Google to rent Google TPUs externally, and Google processed 25 trillion tokens in 72 hours on one model, which tells you Google has the capacity pressure and the incentive to widen distribution past Anthropic. Oracle's entire cloud strategy has been renting other people's scarce compute to customers who can't get it elsewhere, so reselling TPUs fits the pattern it already runs with NVIDIA GPUs. The opposite outcome, Oracle signing the deal but keeping TPUs for internal OpenAI workloads only, is less likely because Oracle makes its money reselling capacity, not hoarding it, and a chip with no NVIDIA tax is exactly the margin story Oracle sells to investors.

Revisit by 2027-06-30: We're right if Oracle offers TPU instances to external cloud customers as a bookable product by its fiscal Q4 2026 earnings call. We're wrong if TPU use at Oracle stays internal or limited to a single named partner with no general availability.