Meta puts prompt-built games straight into Facebook and Instagram
Meta spent Connect 2026 stuffing AI into everything it ships. The Horizon announcement is the one with a distribution business behind it — and a second billion-dollar compute contract landed the same afternoon.
Meta announced Horizon Create, a mobile app, and Horizon Studio, a browser app that build complete 2D or 3D mobile games from a text prompt, including progression systems, difficulty balancing, art direction and multiplayer. Both open in early access through a waitlist. The bigger story is where those games end up: titles published from either tool can be recommended on Facebook and Instagram and played there without a download. Meta's pitch is that someone scrolling a feed can tap a clip and be in a multiplayer session with friends "within seconds," and its blog says games that are "engaging, stable, culturally relevant, and keep players coming back" get more reach.
That last sentence is the real product decision. Horizon never reached scale — it launched as a VR social app and The Verge reported 300,000 monthly users in early 2022 — while Roblox, which now ships a prompt-to-game feature in its own mobile app, reports 123 million daily actives. We covered that launch in Roblox hands creators AI tools — and a path off Roblox.
Meta is not competing on the generator; every platform has one now. It is competing on the storefront, and Instagram and Facebook feeds are a distribution channel no rival can rent. If engagement-gated promotion works, Meta becomes the largest distribution surface for AI-made games without owning a single hit. If it doesn't, this is another Horizon relaunch — and The Verge's Jay Peters is openly skeptical that prompt-built games will be as engaging as handcrafted ones.
Akamai disclosed an $11.6 billion, seven-year commitment from Anthropic, its largest contract ever — with a warrant for up to 5% of Akamai's common stock attached. Akamai filed the deal as an expansion under the master services agreement it signed with Anthropic on May 5, 2026, and says it could grow by up to $9 billion more, a total potential commitment near $20 billion. Notably, the contract covers Anthropic's "accelerating CPU workload demands" through Akamai's distributed cloud — no GPUs are named and no megawatts are specified. The warrant vests in slices: roughly 2% with the $11.6 billion, the rest tied to expansion.
Akamai also raised 2026 capital spending by about $1.7 billion to pre-buy supply chain components, against a projected $5.5 billion in total capital expenditures for the deal. The market reaction was split: shares closed the regular session down 6.78% at $110.41, then jumped about 16% in extended trading once the announcement landed. This is additive to, not a repricing of, the smaller roughly $1.8 billion Akamai deal Bloomberg identified as Anthropic's in May. We tracked Anthropic's wider buying spree here — Anthropic has signed for 14.8 GW of compute — and maybe $517 billion.
The warrant is the tell: to keep winning AI contracts, the infrastructure vendors are now taking equity-shaped risk in their customers.
NYU chemists trained a model to predict where a drug molecule's hydrogen atoms actually sit, and released it open source. The paper, published in Chemical Science by Yingkai Zhang's lab with Xiaolin Pan as first author, mined hydrogen positions from high-resolution crystal structures in the Cambridge Structural Database into more than 1.1 million tautomeric states — orders of magnitude more than earlier experimental datasets — then trained a graph neural network to pick the stable tautomer from 2D structure alone, with no 3D input and no quantum-chemistry run. Applied to 5,075 PDBbind ligands, it flagged 126 cases, about 2.5%, where the assigned tautomer was probably wrong; it also screened the 4.6 million-compound Enamine collection in 3.2 hours on one GPU node.
Why it matters: a misplaced hydrogen quietly corrupts docking, free-energy calculations and molecular dynamics, and most pipelines still infer it. One honest caveat — coverage of this work is currently a single NYU press release, and in the prediction files the team released, the flagship model runs roughly level with the baselines shipped alongside it.
What to watch: whether Meta's feed promotion produces a breakout AI-made game, or whether the waitlist quietly becomes the whole launch.
Would you play a game a friend built from a prompt in your feed — or is a feed of AI-made games the thing you scroll past? Tell us in the comments.
Sources: Tautomer-Predictor (GitHub)
