Tool Equality Is Not Enough

There is a town in Cheshire called Macclesfield, about half an hour south of Manchester. I spent time there in 2013 and 2014, and its centre was hollowing out. Over those two years, the shops that were left kept closing, one after another, and pawnshops opened in their place.

The town had argued about its supermarket before I got there. In 2012, when Tesco planned to replace its Hibel Road store with a Tesco Extra twice the size at Barracks Mill, the town’s own MP warned in public that the new store’s high share of non-food “may suck the lifeblood out of the town centre.” At a public meeting in St Michael’s Church, 157 of the more than 210 people there recorded their opposition, and no one raised a hand in support.¹ People who had lived alongside the first store knew what one twice its size would mean.

For years afterwards I kept asking why the centre had emptied, and the standard explanation did not satisfy me. It holds that the supermarket was cheaper and more convenient, and the shops that could not compete deserved to lose. That is true as far as it goes. It does not explain why the loss felt like more than a set of businesses closing.

The same kind of promise is now made about artificial intelligence. Give everyone the tools the giants have, and the field levels itself. I believed that once, and built a company on it. What I learned is that a market stays open only if small players can enter, reach customers, switch suppliers, and retain enough value to survive. That is what I mean by contestability. Powerful tools guarantee none of it, and it is what the high street lost.

What actually left

A family greengrocer, I eventually realised, is not only a shop. The people who run it are proprietors. They keep their own margin and set their own hours. That income is enough to make them participants in the place they live. They sponsor the junior football side. They are on the committee for the New Year event. The shop is a node in a web of small obligations.

When the superstore replaces the shop, the proprietor’s place goes to shift workers, not to another proprietor. A shift worker is not going to sponsor the football side. They cannot. The commercial transaction still happens. Groceries are still sold. But the surplus that used to flow from the shop into the community stops flowing, because the superstore moves it off the shopfloor and into a corporate margin accounted for somewhere else.

The community surplus was never priced. Nobody ever put a number on the greengrocer’s sponsorship of the football side, so when the shop closed, the loss went unrecorded. Every welfare calculation that says the town is better off is telling the truth about what it measures: cheaper groceries, more choice, more jobs. It is silent about the community surplus.

The scale of the churn was real, whatever you make of its causes. The New Economics Foundation, an advocacy think tank, estimated that Britain lost more than thirty thousand local outlets, from food shops and pubs to bank branches and post offices, in the five years to 2000. Over the same years, by NEF’s account, independent fresh-food specialists saw their sales fall by about forty per cent.² NEF’s 2005 survey found that more than four in ten of the towns surveyed had become “clone towns,” their high streets taken over by the same national chains.³

There is an obvious objection. Nobody forced anyone into the superstore. Shoppers chose it, freely, because it served them better. If the town centre died, the town killed it with its own footfall.

That is true, and at the time it was a good choice. The superstore was cheaper and had everything under one roof, and a chain fighting to establish itself will set its prices for the local market it is trying to take. Britain’s competition regulator documented how. Its 2000 inquiry found the major chains, Tesco among them, persistently selling some frequently purchased products below cost and “price flexing,” which it defined as varying prices in different places “in the light of local competitive conditions,” in ways not related to costs.⁴

But the choices added up, and together they changed the market the next shopper would find. Once the independent shops have gone, they are not coming back, because the footfall that would sustain them now goes to the superstore’s car park. What was a choice among many becomes a choice between the superstore and a longer drive.

The bargain also looks different depending on where you can shop. The large store offers one price; the smaller neighbourhood format offers another. When the consumer group Which? priced a basket of seventy-five groceries, in a comparison published in February 2023, the same shop cost about ten per cent more at a Tesco Express, the small format the chain brought onto high streets, than at a large Tesco: about eight hundred pounds a year more for someone buying that basket every week.⁵ Cheap groceries remain available, but not necessarily on the terms, or in the places, people need them.

The regulator looked twice and acted on neither practice. In 2000 it judged both remedies it considered for below-cost selling disproportionate, and every remedy it considered for price flexing undesirable, disproportionate or impractical. In 2008 it found that below-cost selling did not harm competition, so no remedy arose for it. It did act on weak local competition, ordering the largest chains to give up land and restrictive covenants that kept rivals out, and it recommended a competition test for new large stores.⁶

So “the shoppers chose it” is true, and it leaves out what the town could still choose afterwards. A shopper who can walk to three greengrocers can walk away from any one of them. A shopper whose alternative is a longer drive has a weaker hand, even when the price on the shelf looks fine.

The remedy I built

I did not just watch this. I tried to fix it.

The company was called Cloudfeet, and the theory behind it was simple enough to fit on one line. The reason the greengrocer and the small manufacturer lost was that the giants had tools they could not match. A Tesco or a Walmart runs on planning, forecasting and order-management systems that let it see demand across thousands of stores and move stock against it. The greengrocer has nothing comparable. It orders on what sold last week and on the owner’s judgement. The small manufacturer is in the same position. Selling into a fragmented market, it has no such nervous system either, and its ordinary ERP cannot even model the problem, because it was built to serve a handful of large accounts, not to reach millions of small customers directly. So I built a cloud platform to close that gap. It was a horizontally scalable back end, designed to give the small player the same planning capability the incumbent had and to let a manufacturer coordinate all the way out to a mass of end customers its own systems could never have handled.

Give the small players the incumbents’ tools, the thinking went, and you restore the contest.

It never reached the market. The immediate reason is unglamorous: I could not get investors over the line, and without capital the platform could not reach the scale the model needed. But that reason hides a deeper one that took me years to admit. Even if the funding had come, I am no longer sure it would have restored the contest, because the tools were never what decided whether a small manufacturer could enter a market, reach its customers, or keep its margin.

What the tools did not touch

Four things decided whether a small player could enter a market and reach its customers. The first was capital: the small manufacturer needed working capital to hold stock and absorb late payment before it could enter a new market at all, and no planning system supplies that. The second was logistics: reaching customers meant physical movement of goods, trucks and warehouses and last-mile delivery, none of which a forecast performs. The third was habit: the customer had already been trained, over years of convenience, to reach for the incumbent by reflex, and no back end changes that. And the fourth was distribution: the layer that decides what reaches the shelf, and on whose terms.

I could build the tools. I had no way around the conditions on which they would have to compete.

The bazaar and what sits behind it

Contestability is easiest to see where it is still physical. Walk through Chandni Chowk in Delhi, or Chatuchak in Bangkok. There are thousands of sellers, stall after stall, and the defining feature of the place is that you can walk away. If one trader will not meet your price, the next one will, or the one after that. Here competition is something you experience with your feet. The seller knows you can leave, and that knowledge is the whole discipline.

Now follow the goods backward, off the market floor to the suppliers beneath it. A market can have thousands of sellers and still depend on a much smaller number of suppliers, distributors, or landlords. The visible competition does not tell you how much freedom those sellers have. The shopper on the floor can always take three more steps. The trader behind the counter, looking for a different supplier, a better wholesale price or a cheaper pitch, may find far fewer places to walk to.

In 2005 the European Commission made a set of commitments from Coca-Cola legally binding in the countries of the EU, Iceland and Norway where it had a particularly strong market position. Coca-Cola undertook to stop signing shops and pubs to exclusive deals, to stop paying rebates that rewarded them for buying as much Coke as before, or more, and to stop making a shop that wanted Coke take its less popular brands as well. Where a shop had no other chilled space, it was to be free to use at least a fifth of the Coca-Cola cooler for any drink it chose.⁷ Every one of those commitments concerns the shelf, the fridge and the terms on which a rival reaches the customer.

So the number of stalls tells you little on its own. The question the bazaar puts is whether, somewhere beneath the stalls, the sellers themselves can still walk away.

The same remedy, offered again

The case for open-weight AI is the case I made for Cloudfeet. It offers the same remedy, resting on the same assumption: that better tools are enough. Open models, the argument goes, hand the small player the same capability the giants have. Anyone can download a near-frontier model, run it, build on it, and compete. The equipment has been levelled, so the field will level itself.

And the capability has arrived. Epoch AI’s 2024 benchmark analysis put the lag between the best open-weight models and the best closed ones at about a year; on its newer capability index, a different measure, it has averaged three to four months since 2023.⁸ That is a real gain in what a small firm can make. The app layer that sits on top of these models looks like a bazaar: thousands of startups, stall after stall, apparent competition everywhere you look. You can walk away from any one of them.

But follow the goods backward, the way you would in Chandni Chowk, and the walking gets harder the further down you go. The thousands of app-layer companies sit on a small number of foundation-model providers. By the estimate of Menlo Ventures, an investor in Anthropic, three of them take close to nine-tenths of what US enterprises spend on large language model APIs.⁹

Those providers in turn sit on compute. The chips come mostly from one vendor, Nvidia, with about seven-tenths of the AI chip market by one analyst’s estimate, a more concentrated position than any model provider holds. Beneath that sit three clouds, Amazon, Microsoft and Google, with 28, 20 and 15 per cent of cloud infrastructure spending.¹⁰ Open weights hand you the model file. They do not hand you the compute to train or serve it at scale, and a builder buys that compute from one of a few sellers, on their terms.

Above the builders sits another scarce resource: access to customers. Sometimes a purchasing office decides who gets it; sometimes reputation does, or a platform’s ranking. More capacity to produce does not automatically buy more access. When anyone can build, everyone builds, and the enterprise buyer still looks only at the short list it already has. The app store still ranks what it shows. Whoever controls those allocations holds what the shelf and the cooler held in the Coca-Cola case: the terms on which the goods reach anyone.

Trust is a separate problem from distribution. Distribution decides whether a buyer sees you at all. Trust is whether, having seen you, they will rely on you, and AI has left it roughly where it was. Making things, the writing and coding and designing, has sped up dramatically. Selling them has not, because a sale needs a buyer to decide whether to depend on you, and people still make up their minds at human speed. A relationship still matures over quarters. A procurement cycle still runs its months. So every open-weight builder now produces at machine speed into a market that absorbs at human speed. The constraints I came to see at Cloudfeet, capital, logistics, habit and distribution, have not moved. Open weights loosen the constraint on what a small firm can build, while leaving the constraints on market access intact.

Anthropic’s API tokens do not look as if they are being sold below cost: SemiAnalysis estimates that the company raised its gross margin on inference from 38 per cent to over 70 per cent in the first months of 2026.¹¹ If there is a subsidy, it more likely sits in the flat-rate consumer subscription, where a heavy user can consume far more compute than the fixed fee covers.

Behind that sits the scale of the capital being spent to build the position, with the five largest hyperscalers, Microsoft, Amazon, Alphabet, Meta and Oracle, putting more than four hundred billion dollars into capital investment in 2025.¹² That is spending on acquisition. What the terms look like once the habit is set is not yet known, and it will depend on whether users and builders can still walk away. For the case where they cannot, Lina Khan, then chair of the US Federal Trade Commission, described the “too big to care” posture of monopolies, and Cory Doctorow took the phrase up for tech: big enough to face no consequence for degrading what they sell, because their customers have nowhere else to go.¹³

The disanalogy is real. There are a few such providers, not one, and they compete, sometimes fiercely. So the claim is narrower than monopoly: a handful of providers with high switching costs, at the layer everything else depends on, can leave the plurality above them with much less freedom than it appears to have.

There is also a serious argument that concentration at this layer is a feature, because a small number of well-resourced labs can be held responsible for safety in a way that a free-for-all cannot. But it is an argument about who should hold the power, not about whether the power is concentrated. However the safety question is answered, the builders above that layer still need to be able to enter, reach customers and switch suppliers. A responsible steward does not by itself give them any of that.

The test

When someone tells you a market is competitive, whether it is a high street, a bazaar, or the AI stack, do not count the sellers. The number of sellers is the part you are meant to look at, the stalls and the storefronts and the thousands of startups. Ask instead whether the participants can credibly walk away: whether a seller can change supplier, whether a builder can change provider, whether a small firm can reach a customer without someone else’s permission. Find the layer where alternatives cease to be credible. Look in both directions: below the sellers, at the layer that supplies them, and above them, at the layer that allocates access to customers.

I built Cloudfeet because I thought the small player needed the giant’s tools. The question I would now put to anyone building with AI is whose permission they still need to make a living from it.

Sources

[1] Macclesfield Express, “New Tesco store is ‘no done deal’, says Macclesfield MP David Rutley” (7 March 2012), now at macclesfield-live.co.uk/news/local-news/new-tesco-store-no-done-2526622. Rutley: “the high percentage of non-food may suck the lifeblood out of the town centre.” The 2012 plan was to “replace the Hibel Road store with a Tesco Extra twice the size at Barracks Mill.” St Michael’s Church meeting: more than 210 attended; in a survey of guests by church members, 157 opposed the plans and no one raised a hand in support.

[2] New Economics Foundation, Ghost Town Britain (December 2002). “Over 30,000 local economic outlets” lost in the five years to 2000 (counting food retailers, general stores, pubs, bank branches and post offices); independent fresh-food specialists’ sales down 40% between 1995 and 2000 (NEF citing a Thames Valley University study). NEF is an advocacy think tank.

[3] New Economics Foundation, Clone Town Britain: The survey results on the bland state of the nation (2005). “Of the towns surveyed 41 per cent are clone towns”: 42 of 103 surveys returned by the public, for towns of 5,000 to 150,000 people (plus 27 London returns). Clone town: a high street that has become “a monochrome strip of global and national chains.” Volunteer survey. Not the 2004 report of the same main title.

[4] Competition Commission, Supermarkets: A report on the supply of groceries from multiple stores in the United Kingdom, Cm 4842 (2000), para 1.6. “Persistently selling some frequently purchased products below cost” by the major chains, Tesco among them, found to operate against the public interest; price flexing defined as “varying prices in different geographical locations in the light of local competitive conditions, such variation not being related to costs.”

[5] Which?, policy and insight article on the cost of convenience-store shopping (9 February 2023), which.co.uk/policy-and-insight/article/the-cost-of-convenience-thatll-be-800-please-which-reveals-the-extra-cost-of-shopping-local-aSbR28c4fejd. The same 75 groceries, at 2022 prices, were 10% dearer at Tesco Express than at larger Tesco stores: an extra £15.73 a week, or £817.91 more over a year, for the basket bought weekly.

[6] Competition Commission (2000), para 1.7: both remedies considered for below-cost selling “would require monitoring and intervention that would be disproportionate” (and prohibitions abroad had not proved effective); the price-flexing remedies considered “are either undesirable, disproportionate or present practical difficulties.” Competition Commission, The supply of groceries in the UK market investigation, final report (30 April 2008), summary para 19: below-cost selling not found to be predatory or to produce an adverse effect on competition. The 2008 report did find an adverse effect from weak local competition. The remedies it imposed were the release of restrictive covenants and measures on landsites in highly concentrated local markets (para 42); the competition test in the planning process was a recommendation to government and the devolved administrations, which the Commission could not itself impose (summary para 4, paras 43 and 11.13).

[7] European Commission press release IP/05/775 (22 June 2005). Commitment decision covering the EU, Iceland and Norway, with the commitments “applicable in those countries where Coca-Cola has a particularly strong market position,” until 31 December 2010: no exclusivity, no target or growth rebates (“rebates that reward its customers purely for purchasing the same amount or more of Coca-Cola’s products than in the past”), no forcing outlets to take less popular products with its stronger brands, and, where no other suitable chilled capacity exists, the outlet is “free to use at least 20% of the cooler provided by Coca-Cola for any product of its choosing.” A commitment decision, not an infringement finding.

[8] Epoch AI, data insight on the open-weight to closed-model gap (29 May 2026), epoch.ai/data-insights/open-closed-eci-gap: “Since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months,” measured on the Epoch Capabilities Index, and “slightly larger than” the average of three months between January 2023 and October 2025 in Epoch’s October 2025 data insight (epoch.ai/data-insights/open-weights-vs-closed-weights-models). Epoch’s 4 November 2024 open-models report (epoch.ai/blog/open-models-report), benchmark-based, put the lag at about one year (90% range 5 to 22 months). That is a benchmark-based measure, not comparable with the ECI figures; the two together do not show the gap narrowing.

[9] Menlo Ventures, “2025: The State of Generative AI in the Enterprise” (9 December 2025). Estimated share of enterprise LLM API spend: Anthropic 40%, OpenAI 27%, Google 21%; “these three companies account for 88% of enterprise LLM API usage.” Survey of 495 US enterprise decision-makers; shares estimated from API usage. Menlo Ventures is an investor in Anthropic.

[10] TrendForce press release, “Strong Demand from CSPs and Sovereign Cloud to Drive Over 20% Growth in AI Server Shipments by 2026” (30 October 2025): “TrendForce estimates that NVIDIA will maintain approximately 70% of the AI chip market share in 2025” (basis not stated). Synergy Research Group, “Q2 Cloud Market Passes $143 Billion; Highest Growth Rate in Eight Years” (30 July 2026): Q2 2026 worldwide shares of cloud infrastructure service revenues (IaaS, PaaS and hosted private cloud) of 28% Amazon, 20% Microsoft, 15% Google.

[11] SemiAnalysis, “AI Value Capture: The Shift To Model Labs” (1 May 2026), public portion: Anthropic’s “gross margins on their inference infrastructure have increased from 38% to over 70%” over “this year”, to 1 May 2026. Analyst estimate; labs’ economics are not disclosed. That API tokens are not sold below cost is an inference from this margin, not a statement in the source. The flat-rate subscription point in the same paragraph is qualitative and unpinned.

[12] Epoch AI, data insight on hyperscaler capex and operating cash flow (16 June 2026), epoch.ai/data-insights/hyperscaler-capex-vs-cash-flow, covering Microsoft, Amazon, Alphabet, Meta and Oracle: aggregate cash capex “on track to overtake operating cash flow in Q3 2026.” Epoch’s quarterly figures sum to roughly $412 billion for 2025 (our sum; Epoch states no annual total).

[13] Cory Doctorow, “Too big to care,” Pluralistic (4 April 2024), pluralistic.net/2024/04/04/teach-me-how-to-shruggie/. Doctorow: “Khan described tech monopolists as ‘too big to care.’” Khan’s own words, on The Daily Show (YouTube oaDTiWaYfcM, at about 2:15), were about monopoly power in general: “that type of too-big-to-care-type approach is really what ends up signaling that a company has monopoly power because they can start mistreating you, but they know you’re stuck.” Doctorow applies it to tech in his “enshittification” argument.