Opera Put Free Inference in the Browser — and Google's Flash Prices Double in January

WebAI is a proposal, not a ratified standard, and the distinction is doing real work in the valuation debate. Opera describes it as a proposed web standard backed by companies including Google and Microsoft, and its implementation as model-agnostic and built around open models. So this is less a new capability than a new place to buy one — the browser, rather than the cloud, becomes where inference is sold.

Chrome and Edge were already there. Chrome's Prompt API prompts Gemini Nano on-device through a developer preview, and Edge's Prompt API defaults to Phi-4-mini with Aion-1.0-Instruct available in prerelease — Microsoft has been testing those interfaces in Edge since May 2025. WebAI's genuine addition is routing: a site asks the browser for an answer without knowing whether Gemma, Phi or an Apple model is resident, and without naming a vendor in its own code.

The standardization push is real but unfinished. W3C's \"State of the Web in the AI Era\" report, finalized on October 20, 2026, treats on-device inference as a core web concern and points to the Web & AI Interest Group for follow-up, with a session scheduled for October 29. Until a working group drafts a specification, WebAI is a proposal with three shipping implementations — and two of them existed before October 27.

The dated price signal is not a cut. On the Gemini Developer API, Gemini 3.1 Flash-Lite lists at $0.25 per million input tokens and $1.50 per million output tokens. Gemini 3.8 Flash lists at $0.75 and $3.75 through December 31, 2026, and at $1.50 and $7.50 from January 1, 2027. Gemini 4 Argon doubles on December 1, 2026, from $2 and $10 to $4 and $20. Cloud's cheapest useful tier is scheduled to get more expensive in the same window that local inference becomes free.

Scale decides whether this matters. Opera served 288 million average monthly users in the second quarter of 2026 on $178.1 million of revenue and $42.4 million of adjusted EBITDA. Illustrating ten AI calls per user per day — 60,000 input and 15,000 output tokens each — that workload would cost roughly $10.8 million a month on the cheapest Gemini tier, about $16.0 million on 3.5 Flash-Lite, and $58.3 million a month once 3.8 Flash reprices in January, against monthly revenue of $59.4 million. At browser scale, the cloud inference bill for a free AI feature is the size of the profit line, which is why the company pushing the standard has a motive to. The addressable base today is smaller still: the preview runs on desktop browsers only.

Token prices are where labs defend revenue; per-seat prices are where software vendors defend margin, and per-seat prices are already moving. Microsoft cut the Microsoft 365 Copilot Business add-on to $18 per user per month on annual billing, down from $21, in the same window in which Edge was handing developers free on-device models. Opera's own framing is blunter: users do not need AI subscriptions for WebAI to work.

The inversion is the point. A per-seat vendor sells the same task — summarize the contract, classify the ticket, rank the candidate — that a browser can now do locally for nothing. The vendor that loses is not the one with the weaker model; it is the one whose customer can route the routine half of the work to a free local model and keep the seat for the hard half. That is a discount on the seat, not on the token.

The NPU demand story does not survive contact with the PC market. IDC expects global PC shipments to fall 11.3% in 2026 while PC revenues rise 1.6% on higher average selling prices, with the fourth quarter down 20% year over year and no meaningful memory relief before the end of 2027. A standard that shifts inference to the user's machine raises the compute content of each PC; it does not create PCs.

The clearest beneficiary is not a chip vendor but a memory vendor. Running a local model is not free: Chrome's Prompt API wants more than 4 GB of VRAM, 16 GB of RAM and 22 GB of free storage, and Edge's wants 5.5 GB of VRAM and 20 GB of storage. When Opera shipped local models to developers in April 2024, each variant occupied 2 GB to 10 GB of local storage; the shipping browser requirement two years later is two to three times that. Every device that adopts browser-resident inference carries a model resident in memory.

Cloud buildouts are pulling the same components into servers: Microsoft spent $115.9 billion on property and equipment in fiscal 2026 against $64.6 billion a year earlier, taking capital intensity from 22.9% to 34.9% of revenue, and Alphabet spent $80.6 billion in the first half of 2026 against $39.6 billion. The same shortage is now destroying demand at the other end — Qualcomm warned that memory supply constraints and price increases will hurt handset OEM demand — and lifting memory makers' margins to records. Local inference lengthens the memory cycle rather than ending it, because it adds gigabytes per device into a market that is short of both units and memory.

Agent commerce is where the metered cost of inference actually bites, and the distribution layer is being built right now. Shopify opened checkout to browser-based agents on September 28, 2026 through WebMCP tools that let an agent navigate to a storefront, read and update the checkout, and submit it after the buyer confirms. An agent that completes a purchase is not a demo; it is a bill.

That leaves Cloudflare with a genuinely two-sided case. Revenue grew 36% year over year to $696.1 million in the June quarter, and agent traffic is exactly the kind of load Browser Run monetizes; a $170.0 million quarterly net loss, against $2.5 billion of trailing revenue, leaves little room to subsidize it. But if agents run inside the shopper's own browser with a local model, the hosted browser becomes optional, and Cloudflare is the most exposed name in the stack to that substitution.

Two dates do most of the work in the next quarter. On December 1, 2026 and January 1, 2027, Google raises list prices on the models that sit closest to what a small local model can do. On November 17, 2026, Opera reports third-quarter revenue and adjusted EBITDA and gives its first read on what a browser AI business costs when the inference is free. Opera still discloses no separate AI-inference cost line, which is precisely why that print matters.

WebAI matters because of where the money lands, not because of what it does. A browser that hosts inference at the user's expense takes the marginal token out of a metered API and puts it into a hardware purchase that has already happened — and it does so for the model labs, the hyperscalers and the per-seat software vendors that have been charging by the token or by the seat. The money relocates into memory, into silicon mix, and into the browser that owns distribution. For Opera, that is a strategy the market is barely paying for; for Alphabet and Microsoft, whose cloud revenue is still compounding at 82% and 41% respectively, it is a pricing model that has to be reinvented before January's price step lands.