Finance · Published

Good enough at home? What local AI could mean for OpenAI, Anthropic and the companies behind them

If AI you can run on your own computer becomes good enough for most work, the cloud AI labs could lose customers. Here is what the filings, the usage data and the companies themselves say, in the week Anthropic plans to list its shares.

This is general information, not financial advice. It does not take your circumstances into account. The value of investments can fall as well as rise and you may get back less than you put in. If you are unsure, speak to an adviser regulated where you live.

The question is simple. If an AI model you can run on your own computer becomes good enough for most work, why keep paying OpenAI or Anthropic for every question? This article sets out what the companies' own filings, the usage data and the companies themselves say. It does not say which company's shares to own, and it makes no forecast of its own.

The week it matters

Anthropic has filed to list its shares on the Nasdaq Global Select Market under the symbol "ANT". Its latest registration statement, filed on 10 November 2026, sets a price range "between $70.00 and $78.00" per share. Its IPO website, filed with the regulator, says the price is expected to be set on 18 November and trading to begin on 19 November, subject to regulatory clearance. At the time of writing the registration had not become effective and there was no final price. The filing states no valuation. OpenAI says it has submitted a confidential registration but has not decided on timing.

Why this is not just about two private companies

Some of the world's largest listed companies are tied to the AI labs:

  • Microsoft owns "an approximate 24% interest" in OpenAI on an as-converted basis. It recorded revenue from commercial arrangements with OpenAI of $24.1bn in its year to June 2026, and $10.0bn in July to September 2026 alone.
  • Amazon invested $8.0bn in Anthropic convertible notes from 2023 to 2025 and a further $10.0bn in Anthropic shares in 2026. Its filing says Amazon Web Services and Anthropic expanded their multi-year commitment "by more than $100.0 billion over 10.0 years". Amazon has also invested $50.0bn in OpenAI.
  • NVIDIA says "one direct customer represented 16% of total revenue" in its latest quarter, and that "one AI research and deployment company contributed a meaningful amount of our revenue" through cloud providers.
  • Chip makers are buying in. Alongside the IPO, AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC have agreed to buy "an aggregate of $26.0 billion" of Anthropic shares at the offering price.

If customers left the labs, these companies would feel it.

The case that home AI will be good enough

  • The hardware is arriving. Microsoft's minimum for a Copilot+ PC is 16GB of memory, enough for small models: OpenAI says its open gpt-oss-20b model can "run within 16GB of memory". High-end desk machines now hold far more. Apple's Mac Studio takes up to 512GB (M5 Ultra, from $5,499). NVIDIA's RTX Spark PCs, from $2,599, offer "up to 128GB of unified memory". AMD says its newest processors are "capable of running 300 billion parameter models locally". NVIDIA says its DGX Station supports "large AI models up to 1 trillion parameters".
  • The big companies say so. Microsoft's chief executive, Satya Nadella, on 28 October 2026: "With these new devices, you can run last year's frontier models on your laptop, and last quarter's models at your desk!" Apple's finance chief described Gap Inc. "running AI entirely on Mac Mini ... eliminating cloud token costs". HP's interim chief executive: "the economics alone makes Edge AI compelling."
  • The labs name the risk themselves. Anthropic's prospectus warns that open-weight providers "compete for a growing range of enterprise workloads that prioritize cost efficiency over intelligence", and that customers "may migrate toward smaller, cheaper, or open-source alternatives that adequately address their needs." OpenAI says open models compete "not at the frontier, but on cost, control, and latency."
  • The newest models cost more again. Anthropic's Claude 3 Opus launched at $15 per million input tokens in 2024 and Claude Opus 4.5 at $5 in 2025; its newest model, Claude Fable 5.2, costs $8. OpenAI's GPT-4 launched at $30 in 2023; its newest, GPT-6 Astra, costs $10. Google's Gemini 4 Carbon doubles from $2 to $4 on 1 December 2026. Rising prices at the top give cost-conscious users a reason to look elsewhere.

The case that it will not

  • The newest models are not on your desk. The most capable models from OpenAI, Anthropic and Google cannot be downloaded at all, and the largest open models need data-centre hardware. Even Nadella's own phrase, "last year's frontier models", describes a gap.
  • Home hardware is getting more expensive. Data centres are buying the same memory chips first. Intel expects the 2026 PC market to shrink "approximately 12%". AMD's chief executive said "higher memory and component costs weighed on demand" and that AMD plans for PC shipments "to decline by a double-digit percentage" in 2027. Memory maker Micron's revenue rose from $37.4bn to $133.2bn in a year, mostly from data-centre customers.
  • The labs are still growing fast. Anthropic's revenue was $33.2bn in January to September 2026, against $2.5bn a year earlier, and it had "more than 11,900 customers with $100,000 or more of revenue" over the previous twelve months. OpenAI said in March 2026 that it was "generating $2B in revenue per month", and in September that its products reach "more than one billion weekly active users and 2.5 million businesses". (Both labs are private until Anthropic lists; OpenAI's figures are its own, unaudited.)

What the usage data shows so far

Open-weight models are winning a large share of the work, but not of the money:

Open-weight models: share of the work against share of the money

  • Share of tokens
  • Share of spend
Open-weight models: share of the work against share of the moneyOpen-weight models: share of the work against share of the money. Share of tokens: OpenRouter 72, Vercel 71.5, Menlo survey 19; Share of spend: OpenRouter 27, Vercel 23, Menlo survey 4020406080%OpenRouterVercelMenlo surveyShare of tokens, OpenRouter: 72 %Share of tokens, Vercel: 71.5 %Share of tokens, Menlo survey: 19 %Share of spend, OpenRouter: 27 %Share of spend, Vercel: 23 %Share of spend, Menlo survey: 4 %
Show the numbers
Values (%)Share of tokensShare of spend
OpenRouter7227
Vercel71.523
Menlo survey194
Periods: OpenRouter mid-September 2026; Vercel AI Gateway September 2026; Menlo enterprise survey, the year to September 2026. OpenRouter and Vercel are platforms used mainly by developers and start-ups and measure their own traffic; Menlo is a survey of about 400 US enterprise decision-makers. OpenRouter's spend figure is 100% minus the 73% it reports for closed models. None of the three measures AI run on people's own computers. Source: OpenRouter, Inferred issue 01: The Token State of Play (22 September 2026), checked .
Show all 3 sources
  • On OpenRouter, a marketplace used mainly by developers, open-weight models passed half of all tokens in 2026, and closed models' share of spend fell from 89% in January to 73% in mid-September.
  • Vercel, another developer platform, says: "Open-weight models now process more than seven in ten gateway tokens and account for nearly one in four dollars."
  • In Menlo Ventures' survey of about 400 US enterprise decision-makers, "Anthropic and OpenAI now account for 86% of enterprise LLM API spend, up from 67% last year." Open-source models' share of tokens rose from 11% to 19%, "while its share of spending fell from 6% to 4%, driven by a 3.3x drop in average token prices."
  • Microsoft says "revenue from open source models on Foundry increased 10X" - on its own cloud.

Two cautions. Most open-weight tokens on these platforms are processed in the cloud, often on models from Chinese labs, not on people's own computers. And no source we found measures how much AI actually runs on home and office machines. So far, the evidence shows cheap, high-volume work moving to open models while the paying customers stay with the leading labs.

Three ways to make money from AI, side by side

This is not a ranking. Each model depends on different things going right.

The AI labs (OpenAI, Anthropic). They sell access to the most capable models. What helps them: revenue growth like Anthropic's, and big business customers. What could go wrong: the risk they name themselves - customers deciding that cheaper or open models are good enough. Anthropic's net loss for January to September 2026 was $104.0bn, mostly non-cash changes in the value of its convertible notes and similar instruments as its own valuation rose, and its five largest customers provided about 25% of revenue.

The hardware makers (NVIDIA, AMD, Micron, Apple). They sell to both sides - data centres and desks. In their latest quarters, data-centre sales were the larger business at all three chip makers below. If AI moved onto people's own machines, those machines would need more memory and more capable chips; if it stays in the cloud, the data centres keep buying. The risk is concentration: NVIDIA's single largest customer was 16% of its revenue.

Data centre against devices, latest reported quarter

  • Data centre
  • Devices
Data centre against devices, latest reported quarterData centre against devices, latest reported quarter. Data centre: NVIDIA 89, Micron 34.3, AMD 8.5; Devices: NVIDIA 7.2, Micron 13.1, AMD 3.7020406080100$bnNVIDIAMicronAMDData centre, NVIDIA: 89 $bnData centre, Micron: 34.3 $bnData centre, AMD: 8.5 $bnDevices, NVIDIA: 7.2 $bnDevices, Micron: 13.1 $bnDevices, AMD: 3.7 $bn
Show the numbers
Values ($bn)Data centreDevices
NVIDIA897.2
Micron34.313.1
AMD8.53.7
Periods differ: NVIDIA May to July 2026 (Data Center against Edge Computing, which covers PCs, consoles, workstations, robotics and cars); Micron June to early September 2026 (Cloud Memory plus Core Data Center units against Mobile and Client; automotive left out); AMD July to September 2026 (Data Center against Client and Gaming). Source: NVIDIA Form 10-Q for the quarter to 26 July 2026 (SEC EDGAR), checked .
Show all 3 sources

The companies that have both (Google, and to a degree Microsoft and Amazon). Google builds its own Gemini models, its own TPU chips and runs one of the largest clouds. Google Cloud's operating income was $11.1bn on $30.7bn of revenue in July to September 2026, and Alphabet has started selling TPU systems to outside customers; its filing says "We have signed a limited number of agreements to supply TPU systems". Google Cloud's backlog stands at "$560.8 billion". The cost is heavy: Alphabet's capital expenditure was $51.4bn in that quarter alone. Its filing also shows a large stake in "one private company" that it does not name.

Cloud revenue, each year

  • Amazon Web Services
  • Google Cloud
Cloud revenue, each yearCloud revenue, each year. Amazon Web Services: 2021 62.2, 2022 80.1, 2023 90.8, 2024 107.6, 2025 128.7; Google Cloud: 2021 19.2, 2022 26.3, 2023 33.1, 2024 43.2, 2025 58.7050100150$bn20212022202320242025Amazon Web Services, 2021: 62.2 $bnAmazon Web Services, 2022: 80.1 $bnAmazon Web Services, 2023: 90.8 $bnAmazon Web Services, 2024: 107.6 $bnAmazon Web Services, 2025: 128.7 $bnGoogle Cloud, 2021: 19.2 $bnGoogle Cloud, 2022: 26.3 $bnGoogle Cloud, 2023: 33.1 $bnGoogle Cloud, 2024: 43.2 $bnGoogle Cloud, 2025: 58.7 $bn
Show the numbers
Values ($bn)Amazon Web ServicesGoogle Cloud
202162.219.2
202280.126.3
202390.833.1
2024107.643.2
2025128.758.7
Segment revenue as reported in each company's annual report. Microsoft is left out because it has redefined its cloud segment three times; its Azure figures are in the text. Source: Amazon Form 10-K for 2025 (SEC EDGAR), checked .
Show all 4 sources

The verdict

This is our reading of the evidence, not a forecast and not a view on any company's shares.

Not yet. Home AI is good enough for a growing share of cheap, high-volume work, but the paying customers are staying with the leading labs, and the computers needed to run it are getting dearer.

  • For home AI: small models now run on an ordinary new PC with 16GB of memory, the largest tech companies say last year's frontier models run on a laptop, and the labs name cheaper and open models as a risk in their own filings. The threat is real, and the labs know it.
  • Against it, for now: open-weight models take a rising share of the tokens but a small share of the money. In Menlo's survey their share of enterprise spending fell, from 6% to 4%, while Anthropic and OpenAI rose to 86%. On the platforms that measure it, most open-model work runs in the cloud, not on people's own machines. And the memory that home AI needs is being bought first by data centres, pushing PC prices up and sales down.

What would change this reading:

  • Open models' share of spending, not just tokens, rising for several quarters in a row.
  • A real measurement of how much AI runs on people's own computers. None exists yet.
  • Memory prices falling back, so that machines able to run larger models become cheap.

What to watch

These are the facts to check over the coming months, not predictions:

  • Anthropic's final prospectus and price, expected around 18 November 2026, and its first results as a listed company.
  • NVIDIA's results on 17 November 2026: Data Center against Edge Computing.
  • Whether open-weight models' share of SPEND, not just tokens, keeps rising on OpenRouter, Vercel and in the next Menlo survey.
  • Memory and PC prices, and the share of AI PCs in HP's and others' sales.
  • Any measurement of AI actually run on people's own computers - the missing number in this whole debate.

What these numbers do and do not show

  • Platforms are not the market. OpenRouter and Vercel measure their own, developer-heavy traffic. Menlo is a survey. Consumer apps such as ChatGPT and Claude are not in any of them.
  • Company claims. Hardware makers' model-size claims usually assume compressed models. OpenAI's figures are unaudited company statements.
  • Different periods and definitions. NVIDIA, AMD and Micron report on different calendars; NVIDIA's "Edge Computing" includes PCs, consoles, robotics and cars.
  • The IPO. Price and dates may change until the final prospectus is filed. Nothing here is a view on whether to buy shares in the offering or in any company named.

Disclosure: this article was drafted with help from Claude, an AI model made by Anthropic, which is one of the companies discussed above. Anthropic has no part in what we publish; every fact here is taken from the filings and statements listed under Sources.

Drafted with help from an AI tool, then checked and edited by a person. Sources are listed below.

Sources (37)

Show all 37 sources