The Future of AI Runs on Your Laptop Not in a Data Centre

ByPriscilla Li
A laptop on a desk at night, its screen glowing, running an AI model locally

Novaknown’s forecast: renting AI from a lab’s data centre will end the way the 1960s computer utility did, losing to machines people own. Paid plans will keep delivering less for the same money, the computer on your desk will become good enough for most of the work, and the data centres built for today’s investor-funded rush will sit under-used.

For most people the switch already makes sense. A free app on a laptop with 32GB of memory runs a model that matches what the paid chatbots did a year ago, sends nothing to a data centre and answers to nobody’s lawyers. Coding is the exception, for now.

The squeeze has started. OpenAI lost $9.3bn at operating level on $5.7bn of revenue in the first quarter of 2026, and its investors covered the gap. ChatGPT Pro still costs $200 a month but buys half the usage it used to. Meanwhile the best open-weight models, the ones anyone can download and run, trail the closed frontier by about four months, and the ones small enough for a laptop by six to twelve. A 128GB M5 Max MacBook Pro runs OpenAI’s largest open model at up to 79 tokens a second, at roughly the level the paid frontier reached at the end of 2024.

In 1961, at MIT’s centennial, John McCarthy made the original pitch: computing “may someday be organized as a public utility just as the telephone system is a public utility.” He spelled out the business model:

“Each subscriber needs to pay only for the capacity he actually uses.”, John McCarthy, 1961, as quoted by Simson Garfinkel

People built it. Starting in 1965, MIT’s Project MAC, Bell Labs and General Electric worked on Multics as a “prototype computer utility”. Bell Labs pulled out in April 1969, and GE sold its computer business to Honeywell in 1970. Honeywell cancelled Multics in 1985, and the last site switched off in October 2000. The idea lost to minicomputers, which made computers “affordable to almost all companies”. Once a company could own the machine, it stopped renting time on someone else’s. Rental only came back in the late 1990s, rebranded as the cloud.

A 1960s mainframe computer room with large cabinets and operators at consoles

Today’s labs are making McCarthy’s pitch with a new noun: rent intelligence from our data centre and pay for what you use. This time the landlord charges less than it costs to build what it rents.

$9.3bn lost in a quarter, half the Pro allowance

Sam Altman admitted part of it on 5 January 2025:

“insane thing: we are currently losing money on openai pro subscriptions! people use it much more than we expected.”, Sam Altman, OpenAI CEO, via TechCrunch

He also said he had picked the $200 price himself.

Most of OpenAI’s loss is not the cost of answering chats. Its gross margin rose from 33% to 39% in Q1 2026, so serving customers cost roughly $3.5bn of the $5.7bn they paid. The $9.3bn operating loss comes from about $11.5bn of everything else: training the next model, research, staff, and more than $2.3bn of stock compensation. Heavy flat-rate users are the part that does not even cover its serving costs, which is what Altman described, and documents leaked in 2025 suggested inference was eating revenue in 2024-25. Either way, the subscription does not pay for building the model it sells. Investors do.

OpenAI burned about $3.7bn in cash in Q1 and holds more than $73bn, so it is not running out this year. In Q2 it booked $6.7bn of revenue against a $12.3bn operating loss, and the Financial Times reported a July presentation projecting $278bn of negative free cash flow through 2030.

Anthropic’s first-half 2026 revenue reached $16.3bn, up from $1.1bn a year earlier. Its net loss topped $80bn, but $72bn of that was accounting charges on convertible notes, so most of it was never cash. The cash spending is the part to watch: infrastructure spending nearly quadrupled to almost $9bn, and the company said it could spend up to $750bn on compute in the coming years.

The providers have not raised consumer list prices. They have cut what each plan delivers:

Plan What changed
ChatGPT Pro, $200 10x the Plus allowance, down from 20x; new signups paused in September
Claude Pro and Max, $20-$200 Weekly caps from 28 Aug 2025 on top of five-hour limits
Gemini, AI Ultra now $100 and $200 Compute-based five-hour and weekly limits from 20 May 2026, a significant drop on Google’s own support page
ChatGPT free and Go Ads since 9 February 2026

OpenAI’s Thibault Sottiaux blamed the Pro cut on the compute needs of new DevDay features, and new Pro signups were paused in early September over demand for GPT-6 Astra. That October, OpenAI cut API rates by 50% for GPT-6 Sol and Luna. Compute too scarce for Pro subscribers was apparently plentiful enough to sell to developers at half price. The difference is the meter: an API customer pays for every token, while a Pro subscriber pays $200 however much they use. The squeeze lands on the flat fee, exactly where Altman said the losses were.

Anthropic said its caps targeted people running Claude Code “continuously in the background, 24/7,” and would touch fewer than 5% of subscribers. Karl Kahn’s federal class action, filed on 15 June 2026, alleges Max usage is “far below the advertised amount”. In one five-hour session he used 15% of his weekly allowance.

The cheaper tiers are paid for another way. By 5 May, the self-serve minimum spend for ChatGPT ads had dropped from $200,000-250,000 to zero, and advertisers can now target ChatGPT ads by your prompts and your phone number. Turning personalisation off does not stop it, because ChatGPT’s own settings say “Ads may still be based on your current chat.” The only independent audit, presented at AIES 2026, found more ads served to accounts in lower-income ZIP codes. The people least able to pay for an ad-free tier are the ones whose chats feed the ad system most.

Ads cannot close the gap, which is why the squeeze has to continue. They passed $100m annualised within six weeks. OpenAI’s Q2 operating loss was about 120 times that. Researcher Zoë Hitzig resigned the day the ads launched and told SiliconANGLE:

“the company is building an economic engine that creates strong incentives to override its own rules.”, Zoë Hitzig, former OpenAI researcher

Litigation adds restrictions, and not only for children. After the Raine family sued in August 2025, OpenAI added parental controls. In January 2026 it introduced age prediction that switches on under-18 filters automatically, and adults it misjudges must prove their age with a selfie or ID. Character.AI ended open-ended chat for under-18s on 25 November 2025.

The suits aimed at what adults get are about copyright and defamation. Anthropic’s $1.5bn authors’ settlement won final approval in July 2026, covering about 500,000 works, and the pirated datasets were destroyed. On 24 July 2026 a Delaware court refused to dismiss Robby Starbuck’s defamation suit over what Bard, Gemini and Gemma said about him. On 14 October, eight states sued OpenAI seeking court-ordered changes to ChatGPT. Each suit a provider settles or fails to get dismissed gives it one more reason to refuse, and the refusal lands on the paying adult.

The models themselves are not getting worse. Novaknown’s April review of the Stanford AI Index found benchmark scores still rising. Here, “worse” means caps, ads and refusals.

Your chats are the product

What you type into a mainstream chatbot is kept, read and sold against. Anthropic keeps consumer chats for five years instead of 30 days if users let it train on them. A judge ordered OpenAI to hand over 20 million ChatGPT logs, with names stripped, to the New York Times’ lawyers. Google says human reviewers read a subset of Gemini chats and keep those for up to three years, even after you delete your activity. Since 16 December 2025, Meta has used people’s conversations with its AI to target ads.

A model on your laptop sends none of it anywhere. “We don’t see your prompts or data when you run locally,” Ollama’s documentation says. The guardrails are yours too: the UK AI Security Institute notes that an open model’s restrictions can be removed, so it follows the rules its owner sets, not the ones a provider’s lawyers wrote.

A year behind the frontier, and good enough

Most people don’t need the frontier. In OpenAI’s own study of ChatGPT, practical advice, looking things up and writing made up nearly 80% of conversations; programming was 4.2%. Anthropic serves only 10% of its chat conversations on its top model, against 54% of Claude Code sessions. The frontier earns its price where code and agents run for hours. For everything else, a model from about a year ago does the job, and that is what a laptop runs.

LM Studio is a free app for Mac, Windows and Linux that downloads a model and runs it entirely offline. Qwen3.8-27B, released in August under an open licence, needs about 24GB of memory once compressed. On Alibaba’s own tests it beats Claude Opus 4.6 Max on six of nine benchmarks. Independent measurement points the same way: Epoch AI found that models small enough for one consumer graphics card match the frontier after 6 to 12 months.

The biggest open models are closer still. By Epoch’s capability index they have trailed the closed frontier by about four months since January 2026, and the UK’s AI Security Institute puts the lag at four to seven months, down from six to ten. Those need a server: Kimi K2.6 needs roughly 540-600GB of memory even compressed. On a 128GB laptop the best is gpt-oss-120B, which scores close to OpenAI’s o1 of December 2024.

Apple’s chips are why running it locally is realistic. The M5 puts a Neural Accelerator in every GPU core, with up to four times the M4’s peak GPU compute. Apple’s unified memory is one pool shared by the processor and the graphics chip, so a laptop with enough of it can hold a model too large for most gaming GPUs.

Every chip generation moves the middle of the range. Memory bandwidth, the number that decides how fast a model writes, has more than doubled on Apple’s base laptop chip in five years, from 68GB/s on the M1 to about 153GB/s on the M5. The top laptop chip has gone from 400GB/s to 614GB/s. Novaknown’s own experience matches the numbers: the editor’s previous laptop could barely run a language model, and the current one runs 32-billion-parameter models comfortably.

A 32-billion-parameter model compressed to 4 bits takes about 20GB. Apple’s new M6 MacBook Pro, announced in October, starts at $1,999 with 16GB and can be ordered with 32GB. On an ordinary laptop the model’s design matters more than its size. Qwen 3.6 35B-A3B uses only 3 billion of its 35 billion parameters for each word. In LM Studio on an M4 MacBook Pro, a dense 32-billion-parameter Qwen crawled at under 10 tokens a second while its 30B-A3B sibling ran at about 45, faster than anyone reads.

At the top end, a 16 September benchmark tested a 128GB M5 Max with 614GB/s of memory bandwidth, using the free Ollama runner. gpt-oss-120B generated 67.6 to 79.1 tokens a second, and Qwen 3.6 35B-A3B reached 167.2.

The data centres are making the top end dearer for now. Apple raised Mac prices in June, citing memory costs driven by AI data-centre demand, and a 128GB M5 Max listed at $6,699 in August. The boom is raising the price of its own replacement, with investor money that will run out.

Companies have already started moving. In Cloudera’s survey of 1,500 organisations, 66% had moved AI workloads off public cloud to private cloud or their own servers. Desks follow servers, as they did in the 1970s.

Coders are the exception, for now

For coding, nothing on a laptop matches Fable or Codex today. Anthropic’s Fable 5 scored 80.3% on SWE-bench Pro, a test of fixing real software bugs; Qwen3.8-27B scores 61.7% by Alibaba’s count. The gap is closing fast. Qwen’s April model matched Claude Opus 4.5 on Terminal-Bench, another coding test, about five months after Opus 4.5 launched, and the August model beats Opus 4.6 Max on SWE-bench Pro. At that pace, a laptop model could match today’s Fable within a year.

Data centres full of tokens investors paid for

The boom is real by every measure the industry publishes. Google processes more than 3.2 quadrillion tokens a month, 7x a year earlier. The four biggest spenders plan about $725bn of capex in 2026, up 77%. OpenRouter, a service that routes requests to many model providers, saw weekly volume rise from 0.5 trillion tokens in January 2025 to 128.4 trillion in September 2026. H100 rental list prices are rising again, reversing 2025’s declines: Nebius went from $3.85 to $4.50 per GPU-hour on 1 October.

Rows of half-lit AI server racks with empty aisles between them

Look at what that demand buys, and it follows price. On OpenRouter, open-weight models carry 72% of tokens, still running on rented servers, and in mid-September Chinese labs such as DeepSeek, Moonshot and Z.ai supplied 69%. OpenRouter says these models are popular because they come close to frontier scores “at a fraction of the cost.” Google sells its Flash model on the same budget pressure:

“We’ve heard that many companies are already blowing through their annual token budgets, and it’s only May.”, Sundar Pichai, CEO of Google and Alphabet, I/O 2026

The strongest counter-argument is the order book. Microsoft’s commercial remaining performance obligation reached $627bn on 31 March 2026, meaning revenue already under contract, with a weighted duration of about 2.5 years. A contract is only as good as the buyer, and the biggest buyers of AI compute are the labs. OpenAI holds about $73bn in cash against $278bn of projected negative free cash flow through 2030. Anthropic has floated up to $750bn of compute spending on a business that took in $16.3bn in the first half. Contracts like these pay out only if the buyers keep raising money.

Demand at a subsidised price tells you little about demand at the real one. When investor money stops paying for the models, prices will rise toward full cost. Work a laptop can handle will move to the laptop, and capacity built for the subsidised peak will sit idle. Fortune’s report on Meta’s Hyperion financing raised the precedent of dark fibre, the telecom capacity laid during the dot-com boom that went unused for years. That is Novaknown’s forecast.

The losses won’t land on the companies that booked the capacity. Meta’s Hyperion campus sits in a $27bn vehicle 80% owned by Blue Owl, off Meta’s balance sheet and financed with bonds. The grid is paid for through electricity rates: when a utility builds lines and buys capacity for a data centre, the cost is spread across every customer’s bill unless a tariff assigns it to the data centre. Twenty-three states have set special tariffs for large users to make them carry more of that cost, which tells you it was landing on everyone else. On the PJM grid, which serves 13 states in the eastern US, the data-centre share of costs has been put at $9.3bn, and the NRDC estimates families could pay $70 a month more by 2028. The bill for AI datacenter spending has bondholders and households on it.

What happens next

The labs’ costs are going up and yours are coming down. They sell tokens for less than it costs to build the models behind them, cut what each plan delivers and fill the free tiers with ads, because investor money is the only thing closing the gap. Open models that fit on a laptop trail the frontier by six to twelve months, and every new laptop chip runs a bigger one faster.

When the subsidy ends, rented AI will cost what it really costs. The first test is Anthropic’s November IPO, where some investors are now discussing a valuation of about $1.5tn, not the roughly $2tn they had expected. Most of what people use AI for will move to the machine on their desk, and the data centres built for the subsidised rush will be this boom’s dark fibre.

You don’t have to wait. LM Studio and Ollama are free, and a laptop with 32GB of memory runs Qwen3.8-27B today, or the faster Qwen 3.6 35B-A3B. Once it’s on your desk, nobody can halve your allowance, read your chats or decide what you may ask.

Further Reading