The AI industry has a Detroit problem and consumers will decide how it ends

Western AI's Big Three are making the same mistakes Detroit's Big Three made in the 1970s. Chinese open-weight models are cheaper, getting better, and hard to beat. Where does that leave us users?

by Denis Haman

Published:

In the 1960s and early 1970s, Japanese cars were a bit of a joke in America. Cheap, tinny, underpowered, they were – with much help from US automotive boosters of course – derided as what you bought if you couldn't afford a real car. Detroit, and its mighty Big Three of Ford, GM, and Chrysler, wasn't worried.

Back then, the American car was a cultural institution well beyond mere transport. A big block US auto represented the unstoppable might of US engineering and of ballooning economic power, and you’d have been hard pushed to find anyone who thought a Toyota was going to threaten a Chevy.

But the quality gap closed, and while US rivals were focused on each other and producing larger and more lurid land yachts which raised prices and running costs, Japanese carmakers were carefully improving everything they did, from factory floor to showroom, while driving running costs lower and lower.

And then the 1973 oil crisis struck. Running costs became the top priority for consumers who didn’t have the money to watch the gas needle droop with every mile. As well as bringing in the 55mph speed limit, it made usage economics a yardstick we have yet to put down.

Detroit, in a panic, lobbied Washington hard. The resulting Voluntary Export Restraint – “voluntary”, or face an absurd tariff – capped imports aggressively to give Detroit breathing room, but by then US consumers were pretty keen on the efficient and now hyper-reliable Japanese brands.

Instead of holding the tide back, the VER import cap saw Japanese carmakers pivot to build plusher pricier models and new brands with much higher margins – Lexus, and Acura – as well as building locally in US factories.

Consumers dictate the story, as they always do, and within a couple of decades the Big Three of Detroit, Ford, GM, and Chrysler, were shadows of themselves in the city, the latter two then barely emerging from 2009 bankruptcies propped up by bailouts and takeover capital. Only Ford survived on its own merits, and only just.

The US AI industry is in danger of being Detroited, this time by China. The Big Three of AI - OpenAI, Anthropic, and Google DeepMind - are to most observers unstoppably well-funded and already technologically and culturally dominant in the new superindustry of our era. But there is a reason they are already hyper-focused on lobbying and restrictions rather than the economics and consumer habits they should be.

Chinese “open-weight” AI rivals faced a period of being perceived much as Japanese cars were in the 60s - raw, cheaper, unreliable - while also facing legislative hurdles in the shape of chip export controls, safety arguments, fears over access, and of course naked national interest. But while it took several decades for the net effects of ignoring your rivals and consumer preferences to strip Detroit of its people and power, the US-China AI battle is unfolding in front of our eyes and wallets.

What does it mean to us?

Publishers are not just defending against the effects of AI on their business, they are consumers of it too.

A closed model like Claude or ChatGPT sits on its maker's servers. You use it through their channels, on their terms, at their prices, at costs for enterprise and team users which can make you wince before even entering a single prompt.

An open-weight model puts its code into the world so you can download it, run it, host it cheaply, and walk away without asking permission or ever signing a hefty enterprise contract. And these models are fast, cheap, and sold on terms so simple they feel old-school: use the API and pay per token, or host your own and pay only for power and upkeep.

I've been living this. I use Kiro, Claude, and OpenAI tools every day, and I mean intensive use. But I keep getting locked out of Claude daily by lunchtime, and weekly caps breached by Wednesday. Hitting them doesn’t stop your bill - that keeps going up if you need to keep working - but it halts all progress until you dip into your pocket again. Like a 70s gas guzzler.

This isn't just a me thing: Anthropic has acknowledged throttling at peak hours, Forbes investigated "huge pricing issues with glitching Claude Code limits", and subscribers have been vocal about burning through weekly limits in half the time and being unable to work. There is now a proposed class action over the issue.

How does this effect publishers?

Publishers buying at scale use an API where every unit of work is visible, and while individually it does not look a lot, it can very rapidly mount up.

A token is roughly three-quarters of a word. A typical editorial task - say, read a 600-word article, suggest some headlines, write an SEO description - will use about 2,000 tokens, around 10 cents at top Claude pricing. That‘s fine in isolation, but if you run 500 tasks a day across a publisher you could quickly spend between $15,000 and $50,000 a year on that one model alone. And do we think the current AI giants are planning to slash their prices?

Eyeing the world as a consumer will, DeepSeek V4-Flash, launched in August 2026, currently costs about 3 cents per benchmark test against $3.15 for Claude Fable 5, but it has no session windows or weekly budgets, and doesn’t need a huge onboarding cost for an enterprise team.

The same 500-task daily workload drops from up to $50,000 a year on Claude to under $2,000 on DeepSeek. And lo, this very day, DeepSeek has announced a "significant" price increase is coming, with details to follow. Even the disruptors, it turns out, can't avoid the consequences of their own competitors.

And the quality gap? Across 35 shared benchmarks, Kimi K3 beats the top Claude model on 12 and ties on one, despite being 70% cheaper per token. In fact on cost-adjusted performance, Kimi K3 delivers 2.8 times more solved tasks per dollar.

Recognise the pattern? Near-equivalent output, at a fraction of the price. Detroit, Datsun, 1974 all over again.

One thing to put out here. Glide builds its AI assistant GAIA on AWS Bedrock, so of course I should outline our position here as it is consistent with what people want. Bedrock charges only by token use so has no subscriptions or minimums, nor separate use agreements with any model provider. Per-token rates match official rates exactly, but you get all the enterprise backup you would expect from AWS, and can swap models literally with a couple of clicks. There are over 50 models now, so there’s plenty to choose from.

The infrastructure gap is a myth

Western commentators assumed chip sanctions would contain Chinese AI hardware development, but that assumption is already well out of date.

The specialist chips which power AI models are expensive, Western-controlled, and until recently unavailable to Chinese firms. Echoing the Japanese carmaker move to build better, when Huawei was cut off, it built its own chips. The Ascend 910C already delivers about 60% of the performance of the H100 Nvidia industry standard, and the next aims to more or less match it. Huawei's share of the Chinese domestic AI chip market grew from under 10% in 2021 to over 40% in 2025. Sanctions designed to hobble the firm had the opposite effect and not for the first time in commerce barriers quickly became more like springboards.

The open-weight refuseniks

Ironically, the US firms can already see that stamping down on an open-weight future is likely to be counterproductive. In July Nvidia led a coalition of over 70 companies, including Microsoft, Meta, Hugging Face, and Google, in signing an open letter against restrictions on open-weight AI models. Anthropic didn't sign, while OpenAI added its name only after the letter was already published and it was criticised extensively for its position.

None of this is surprising. Companies whose revenue depends on closed models have an obvious reason to limit open alternatives. But the position became untenable in the face of sheer consumer will. OpenAI signed in the end because staying out brought reputational cost at a time when it needs every bit of goodwill it can get.

The US car industry held its position until it couldn't. The timeline from "not a serious competitor" to "we need government intervention" was eight years from Oil Crisis to import caps; the AI version is moving much faster.

The singularity ramble

OpenAI has been losing ground fast. ChatGPT's market share fell from 87% to around 55% in twelve months, while Anthropic overtook it in US enterprise adoption for the first time, adding fuel to speculation about CEO Sam Altman's position.

When the firm’s own models broke out of a testing environment, and hacked into Hugging Face - curiously, a direct competitor - and accessed four other companies in the process, alarms really started shrieking, albeit not at OpenAI.

15 US state attorneys general have demanded OpenAI preserve all materials related to the Hugging Face attack – including tips and tricks left by the model for future agents on how to escape OpenAI's controls – and legal exposure runs from federal hacking charges to civil litigation.

While it was surely time for Altman to turn statesman and to make clear and unequivocal statements about this accidental AI nuke, he instead appeared on a tech entrepreneur podcast and rambled on about being “like, in the singularity”. Awesome, bro.

Anthropic and now Meta have also revealed their own errant AI hackery, also definitely accidents, so this particular virus of irresponsibility seems to be spreading.

Where this leaves publishers and media

Publishing spent a decade being squeezed by big tech. Now that big tech is being squeezed by Chinese big tech, and the pressure is backing up down the pipe.

The subscription tiers, rate limits, and overage charges are not just occasional annoyances, they are the sort of things which strangle an industry and become structural barriers to consumers. That’s us, in publishing and media. The open-weight alternatives, whatever the geopolitical complications, are making those barriers harder to tolerate every month.

This is not a call to abandon Western AI, but it is a fundamental question for its providers: will you build an ecosystem and business models that work for the businesses that need it, or keep the walls high and watch your customers move to wherever makes the numbers work?

This is a first-time crisis for AI, but publishers have been here before. We’ve navigated platform dependency, advertising collapses, and algorithmic upheavals, and know what it looks like when an industry finds the need to reinvent itself time and again.

For AI, is the foot on the gas pedal, or the brake?