Mark Zuckerberg is committing $135B to a radically completely different AI technique, one which positions Meta because the distributor of superintelligence to people, not enterprises.
Mark Zuckerberg revealed his case for private superintelligence on Monday and, the identical morning, shipped the factor that makes the case reasonably priced.
In The Future is for Everyone, dated 10 August, he reaches for a comparability this publication is called for: the cities that introduced railroads, highways, electrification and broadband grew to become centres of analysis, enterprise and business, and he argues AI infrastructure can do the identical for the locations that host it. The promise on supply is an agent that works “24/7 in your behalf” throughout well being, profession, funds, residence and hobbies, free for billions of individuals, with heavier customers priced by what he calls a “dynamic public sale mechanism.”
I consider everybody ought to have entry to superintelligence, and I wrote a protracted piece about Meta’s philosophy and values for constructing a optimistic future for everybody. https://t.co/2ZoNZXZ39T
— Mark Zuckerberg (@finkd) August 10, 2026
Hours earlier than it went up, Meta Superintelligence Labs revealed Muse Glimmer: a 30-billion-parameter agentic mannequin, weights launched beneath a permissive Apache 2.zero licence, sufficiently small to run on a Mac or PC with a single shopper GPU. That’s the manifesto’s economics, shipped as a checkpoint fairly than a pledge.
What private superintelligence seems to be like when it really works
Zuckerberg’s examples are home and present-tense. His agent flags info and helps him prototype; it displays his sleep, then watches him prepare and provides suggestions; it plans personalised recipes for him and his daughter to bake every weekend and orders the substances. His eight-year-old codes her personal concepts and produces movies in a night, and the 2 of them are designing a robotic collectively. Meta researchers, he writes, are producing novel crystal buildings suited to augmented-reality glasses.
Comply with that line ahead and the claims get bigger. A tutor with a PhD in each topic, giving the assistance that presently goes solely to college students whose mother and father pays for it. Biohub’s open-source virtual-cell and protein fashions already within the fingers of researchers designing medicine, in opposition to a aim Zuckerberg and Priscilla Chan set of curing or stopping all illness this century, which he says he now expects sooner. Occupations that hardly exist but: one-person studios designing customized furnishings and clothes, world builders, private biologists formulating particular person remedies. These are his projections, and they’re labelled as such all through the doc. They’re additionally the clearest assertion but of what the capex is for.
Why price per person decides Meta’s wager
A persistent agent for a billion individuals is a distinct inference downside from an enterprise API. Enterprise pricing recovers price per name. A free agent operating constantly doesn’t, and the promise of round the clock service commits Meta to standing capability at a scale it has to pay for earlier than anybody pays it again.
The stability sheet exhibits the size of the dedication. Meta narrowed its 2026 capital-expenditure guidance to $130–145bn on 29 July, from a previous vary of $125–145bn, and recorded $31.08bn of capex within the second quarter alone. Full-year 2025 capex was $72.22bn. Free money move for the quarter got here in at $784m, in opposition to $8.55bn a 12 months earlier.
Muse Glimmer is the primary concrete reply to that arithmetic, and it factors away from the info centre. At full precision a 30-billion-parameter mannequin wants over 55GB of reminiscence; Meta compressed the weights to roughly 4-bit precision, bringing the language mannequin beneath 20GB and leaving headroom for the KV cache, the image-understanding encoder and a speculative-decoding drafter inside a 24GB or 32GB envelope. Meta says the compression price minimal to no degradation on agentic duties. By itself revealed comparability, Glimmer leads Gemma4-31B and Qwen3.6-27B on MCP Atlas at 75.5 in opposition to 54.2 and 62.5, and scores 51.2 on SWE-Bench Professional, whereas trailing Qwen on computer-use and terminal work. Meta’s benchmarks, on Meta’s chart, awaiting third-party runs.
Regionally. That’s the load-bearing phrase. A mannequin constructed for always-on native workflows that retains working with or with no community connection is an agent whose marginal token Meta by no means pays to serve. Zuckerberg has stated weights for Muse Spark 2.1 comply with within the coming weeks, which suggests that is the course fairly than a one-off.
What the dynamic public sale nonetheless has to show
The manifesto’s pricing declare is that the public sale will assure everybody the bottom attainable value for the intelligence and compute they use, whereas directing capability in direction of no matter individuals collectively discover most useful. That could be a stronger declare than a rationing mechanism, and it arrives with no unit, a reserve value or any description of how the market clears. Publishing these mechanics is the only most helpful factor Meta might do subsequent, as a result of it’s the place an affordability promise both turns into a quantity or stays a sentence.
Two different commitments are value holding Meta to, and each are low-cost to confirm later. The primary is governance: Meta says it’s giving its unbiased board of administrators the facility to approve the security standards for mannequin releases and to assessment whether or not every launch meets them, and encourages different labs to undertake the identical construction. The second is the group compact behind the buildout, of which Zuckerberg’s personal instance is Richland Parish, Louisiana, the place he says academics obtained a $50,000 bonus this 12 months from the tax income of a Meta knowledge centre — funded by means of a $1bn fund directed on the US communities internet hosting its amenities. Meta additionally commits to restoring extra water than it makes use of in its watersheds by 2030, and 200% of it in areas of excessive water stress. These are checkable in twelve months. Checking them is how the compact turns into a typical as a substitute of a paragraph.
The tracks being laid right here run from the info centre to the desk, and now to the telephone. On 10 August a mannequin able to planning, calling instruments and recovering from its personal failures grew to become one thing you’ll be able to obtain and run on {hardware} you already personal. The weights are on Hugging Face. Regardless of the public sale ultimately expenses for the compute that won’t match on a laptop computer, the model that does match is already free, and the following one is just a few weeks out.
Alex Chen Alex Chen Read More
