Meta's $100 billion AMD wager and the inference economy behind agentic assistants

The computer that never closes: what Muse promises

The idea of a computer that keeps working after you pocket your phone sums up the always-on assistant in a single line. A Meta employee who let the Muse app handle a three-week honeymoon itinerary and described it as the trip's third participant shows the product is built as more than a chat window. Launched on September 8, it passed five million downloads across the United States and Canada in about twenty-two days, reaching the mark far faster than ChatGPT at fifty-six days or Grok at over a hundred. These download-race figures are reported by finance.yahoo.com citing Sensor Tower data and comparing the rival assistants side by side.

Every new user receives what Meta calls a secure virtual machine : a small computer running in Meta's own data center, with its own browser, files and memory, sealed off from every other user's agent. Tell it to book a table for Friday and track the price, put the phone away, and that machine keeps clicking, waiting and checking in the background. Before sending an email or paying for anything, it pauses and asks; a separate supervisor on the same machine, named Senator, must approve anything reaching the internet. Turning a saved recipe into a grocery list, remembering dietary restrictions, sending the invitations and then paying through Stripe with a single-use virtual card makes the agentic behavior promise concrete.

The price of thinking: the inference economy

The expensive part of this picture is not the machine but the machine's thinking. A setup with two virtual processors and eight gigabytes of memory matches an ordinary cloud server rentable for roughly forty dollars a month. The real meter runs as the model fires again to plan each step of the assistant; watching a flight price every hour for a week means one hundred sixty-eight checks, and every check wakes the model up. Training is a setup cost paid once, while inference becomes a recurring operating expense for as long as people keep using the system.

Meta prices this expense directly: basic use is free, a power plan costs twenty dollars a month and the top tier one hundred. The company's chief AI officer said openly that these tiers help cover computing costs, which is Meta confirming in plain language that thinking costs money. The scale side is striking: Meta's apps reached a daily average of three point six six billion people in June, and Muse already works inside WhatsApp, so even a small slice of that audience running background agents means hundreds of millions of little offices. The daily-user and quarterly revenue figures appear in the official results release on investor.atmeta.com as three point six zero billion and sixty point eight billion.

Spending runs at matching tempo: capital expenditure is guided at one hundred thirty to one hundred forty-five billion dollars for 2026, with about thirty-one billion spent in the second quarter, up eighty-three percent on the year before. As a result free cash flow fell from around eight and a half billion to under one billion. Subscriptions alone will not pay that bill: in the brokerage scenario cited by the host, one billion Muse users producing roughly ten point eight billion dollars of revenue in 2027 would still equal only seven to eight percent of a single year of infrastructure spending. The upward revision of the spending range is reported by fortune.com as a band of one hundred twenty-five to one hundred forty-five billion. The sharp fall in free cash flow is highlighted by fool.com as seven hundred eighty-four million kept from thirty-one point nine billion of operating cash.

The six-gigawatt wager and the ten-percent hook

The chip-side answer arrived on February 24, when Meta and AMD announced an agreement spanning several generations and covering up to six gigawatts of AMD Instinct graphics processors. One gigawatt roughly equals the output of a large nuclear reactor, so the discussion is about six reactors' worth of AI silicon. The first gigawatt comes as a custom accelerator based on the MI450 architecture, with shipments starting in the second half of 2026. The AMD chief executive spoke of double-digit billions of dollars of value per gigawatt, with the total reported above one hundred billion. This framework is described in the official announcement on about.fb.com as a multi-year, multi-generation roadmap. Shipment timing and Helios architecture details are given in the press release on ir.amd.com as the second half of the year.

The most interesting clause is not the sale itself but the mechanism rewarding it. AMD granted Meta a warrant for up to one hundred sixty million shares, about ten percent of the company at one cent per share. The rights do not arrive at once: the first tranche vests only when the first gigawatt actually ships, and the rest unlock as later gigawatts ship and the share price clears hurdles, with the final tranche conditioned on six hundred dollars. The stock stood at one hundred ninety-seven dollars on signing day and crossed six hundred in September, yet Meta cannot collect most of the shares without keep ordering. Everything beyond the first gigawatt is framework rather than firm order; walking away early leaves most of the equity on the table.

The structure chains two giant customers into partnership, since an almost identical six-gigawatt framework was set up with OpenAI in October 2025. The similar equity mechanism on the OpenAI side is described by reuters.com with a ten-percent tranche and a thirty-four-point jump in the shares. A company reaching billions of users daily, with one of the strongest engineering teams in the world, tying its product roadmap to AMD counts as the strongest evidence for the second-supplier thesis. AMD's data-center business more than doubled to six point seven billion dollars in the June quarter, making up fifty-eight percent of total revenue, and even the low end of one gigawatt exceeds a quarter of that business. In a market where a single supplier sets prices and priorities, big buyers naturally wanting an alternative is no surprise, and the chief executive's comeback against Intel in server processors feeds that confidence.

The counter-thesis arrived in July from a chip research firm. Its finding holds that the Meta-specific MI450 carries roughly twenty-five percent less computing power than the standard version, holds less high-speed memory, and keeps only half the chip-to-chip bandwidth of the rival's newest architecture. The review calls the design weak for frontier AI and tailor-made for the recommendation systems ranking Facebook and Instagram feeds. It further warns that Meta's best AI lab could favor rival chips for its next models. The custom chip's memory and bandwidth gaps are summarized by tomshardware.com based on the research firm's findings.

The risk does not stand alone: although the first gigawatt is a firm order, shipments have slipped into the second half of the year, racks are still in testing, and Meta's separate arrangement with the rival chipmaker continues. The Muse side shows wobbles too: in a pre-launch trial, a ticket-watching agent stopped refreshing the page after about fifteen minutes and silently skipped errors, while the chief technology officer wrote that the app kept logging him out repeatedly. If agents prove flaky, nobody leaves offices running in the background and the expected extra computing demand never materializes. Even so, feed ranking itself is inference work, and Meta running the world's largest recommendation systems is on record in the rival chief executive's own words. Because the pact spans generations and the reward depends on later purchases, Meta has strong motivation to fit these chips in; the signals to watch are slipping first shipments, stopping at one gigawatt, and stumbling agents.