Does customer concentration risk matter in the AI industry?

The article starts from a figure published by Ramp on the degree of revenue concentration at the major AI labs, a figure that allows for two opposite readings. The optimistic reading holds that the top 1% of customers who spend the most today are simply the early-adopter segment, and that the rest of the customer base will converge over time toward similar spending levels. The pessimistic reading, by contrast, sees in that same 1% the customers with the most bargaining power and the strongest incentives to migrate to open models, which would erode the labs' margins. The author uses this contrast as the starting point for a broader question: customer concentration is a phenomenon that recurs in practically every layer of the AI value chain (models, chips, cloud, networking and manufacturing), but concentration is not automatically synonymous with bargaining power for those customers. To demonstrate this, he walks through concrete examples in each layer, rating each company's concentration risk as low, medium or high depending on how vulnerable that company is to its large customers switching suppliers, demanding concessions or cutting spending.

In the model layer, Anthropic is presented as a medium-risk case. Its S-1 (the prospectus filed ahead of an IPO) is expected to be published in the coming weeks and to provide more detail on customer concentration, but in the meantime some figures are known: 80% of its revenue comes from B2B customers, it has 6,000 accounts spending $100,000 or more a year, and it went from 500 to 1,000 accounts spending more than $1 million a year in less than two months, with an estimated net dollar retention (NDR) of 140%. What is not known, however, is what percentage of revenue its ten largest customers represent. One relevant data point is that Meta, rumored to be a major Anthropic customer (SemiAnalysis estimates between 3% and 5% of its revenue), has already reached the technological frontier, which raises a real risk that it will drop Anthropic as a supplier. The analysis notes that we will have to wait and see, once the S-1 is published, what the attach rate is for Anthropic's products beyond coding.

OpenAI, by contrast, is rated low risk. Having postponed its IPO, there is less data available, but it is known that more than 50% of its revenue already comes from the enterprise segment, that this officially overtook the consumer segment earlier than expected, that it had 9 million business seats in February 2026, that Codex reportedly has between 15 and 20 million weekly active users, that its advertising business is already running at a $1 billion annualized rate, and that, according to Ramp, its growth in the third quarter was faster than Anthropic's, albeit from a smaller base. The conclusion is that, compared with Anthropic, OpenAI has a broader product portfolio and a more diversified revenue base.

The article then introduces a reflection on the so-called “Fable 5” (as it is named in the original text), whose limited adoption, together with the growing adoption of open-source models, would suggest that the intelligence threshold needed for most tasks has already been crossed, which threatens the price premium that frontier models have historically enjoyed. Against this, it is argued that Anthropic has the opportunity to expand beyond code and make most knowledge work as “legible” to scaling laws and RLVR as programming has been (which would justify, as rumored, a $30 trillion total addressable market). The author believes the Jevons Paradox will hold as prices per token fall and adoption soars, although it is unclear who will capture that incremental token spending: at the current, relatively stable stage, the enterprise and application-level post-training market is said to be maturing to the point where much of enterprise token spending could be reallocated toward an ecosystem of specialized models with a far lower cost per task than that of the frontier labs' models.

In the chip layer, Nvidia receives a medium-risk rating. The central thesis, citing Ben Thompson, is that as long as the world remains fragmented Nvidia wins; its margins would only come under pressure if a limited number of buyers had the means to buy only part of its technology stack or to replace it entirely. In recent weeks Nvidia has bought Hugging Face and licensed Poolside's model development factory, moves the author reads as an effort to “commoditize its complements”: for Nvidia, a thriving open-source ecosystem is existential in order to avoid depending on just a couple of customers that are investing heavily in replacing it. That is why Nvidia began breaking out its data center revenue into two segments: hyperscale and ACIE (which covers AI clouds, industry and enterprise). ACIE customers, such as the neoclouds, are loyal to Nvidia, while hyperscale customers have the resources and incentives to develop their own silicon. Nvidia's dominance in training is far more solid than in inference: the Jalapeño results demonstrate the viability of custom silicon aimed at inference workloads, and AI-assisted electronic design has already begun to shorten time to tape-out. Google's TPUs and Amazon's Trainium already serve the frontier labs, though this also reflects how constrained compute supply is. Microsoft's Maia and Meta's MTIA are further behind, but remain priorities. The text draws a parallel with the Wintel duopoly and Intel's fall with the arrival of mobile phones, where Arm's instruction set proved superior for low-power computing, and asks whether AI inference could be to Nvidia what mobility was to Intel, though it concludes that the answer depends on how large the gap between training and inference workloads really is compared with the far more drastic gap between PC and mobile.

Cerebras, by contrast, receives a high-risk rating. Before its IPO, G42 was by far its largest customer, and in 2026 OpenAI and G42 will together account for 94% of its revenue, with OpenAI becoming the overwhelming majority in 2027 and 2028. Compared with Nvidia, Cerebras is far more concentrated in every sense: focused more on inference than on training, serving two customers (soon one) with the means and incentives to develop their own silicon, and with no clear evidence of a new customer cohort equivalent to Nvidia's ACIE segment.

Broadcom, on the other hand, gets low risk. Its portfolio is broader, including VMware, and its role is not to design its own chips but to co-design and implement the hyperscalers' custom silicon. Semiconductors account for $15.01 billion, 68% of its revenue (up from 56% a year earlier); infrastructure software (VMware) accounts for $7.18 billion, around 32%. Within semiconductors, AI represents $10.8 billion, 72% of the chip segment and roughly 49% of the entire company; within that AI business, networking accounts for close to 40% (more than $4 billion per quarter) and custom XPUs the other 60%. Broadcom's moat is what the author calls “process power”: physical implementation requires coordinating process rules, HBM memory, packaging, power, signal integrity, networking, validation and production schedules. Its own filings stress that relationships with key customers can span years of joint development and accumulate customer- and system-specific knowledge; a customer can switch implementation partners, but that entails additional engineering work, schedule risk, validation burden and the loss of accumulated learning. Broadcom's core competencies are not core to the hyperscalers' custom silicon efforts.

In the cloud layer, CoreWeave receives high risk: its two largest customers accounted for 62% of its revenue in the second quarter. Its dependence on Microsoft and Meta was underscored when, in the weeks before its IPO, rumors emerged that Microsoft was cutting back some of its commitments, which caused widespread concern among investors; shortly afterward OpenAI announced a $12 billion commitment that calmed nerves. The article introduces a concept that recurs on Microsoft's earnings calls, the “fungibility” of compute: the idea that if capital investment in AI does not generate the expected return, that computing capacity can be reallocated to other workloads. Microsoft is also a customer of its own Azure cloud (which in fact partly slows Azure's reported growth) and prefers to own its data centers rather than lease them, given its experience building them. If compute is largely fungible, this raises questions about the role of neoclouds like CoreWeave, which mostly operate Nvidia fleets. CoreWeave earned the Platinum rating in SemiAnalysis's ClusterMAX neocloud ranking, ahead of Nebius, OCI, Crusoe and Fluidstack, thanks in part to its PaaS offering for inference; even so, compared with other neoclouds, it is far more concentrated and exposed to the hyperscalers' capital allocation decisions.

Nebius receives medium risk. Like CoreWeave, it counts Microsoft and Meta as customers, but it also has companies such as Shopify and Cloudflare. Thanks to its origins in Yandex, it has even tighter optimization across in-house designed servers and racks, and the article notes that how Microsoft's and Meta's weight in its customer mix evolves will be key to its trajectory.

In networking, Lumentum receives medium risk. It is a key player in optical networking, supplying lasers for optical transceivers and co-packaged optics; Google accounts for around 26% of its revenue, with Microsoft and Amazon contributing between 8% and 15% each, and Nvidia, Cisco and Ciena also among its customers. Like Broadcom, its moat is partly “process power”: manufacturing advanced lasers requires specialized semiconductor facilities, accumulated manufacturing know-how and consistently high yields; designing your own accelerator does not automatically give you the ability to manufacture its optical components. Lumentum has between 50% and 60% share in cutting-edge 200G EML lasers, and its components go through reliability testing that stretches over several quarters. It also has exposure to competing network architectures: both conventional pluggable transceivers and emerging co-packaged optics require lasers, so the architecture can change without eliminating the need for its underlying capabilities. The most immediate risk for Lumentum is therefore spending, not substitution: Google does not need to learn how to make lasers to hurt its growth; it need only delay deployments. Lumentum's case illustrates the distinction between customer concentration and those customers' bargaining power: its buyers are concentrated, but so are the capabilities they need.

Arista also receives medium risk. Microsoft and Meta accounted for 42% of its revenue in 2025 (26% and 16% respectively), but, unlike Cerebras, Arista has a credible path to adding another large customer: Anthropic is the clearest candidate to exceed 10% of its revenue, since it uses Nvidia GPUs, Google TPUs and Amazon Trainium, and operating across multiple platforms generates demand for a common networking approach. Arista's moat is not Ethernet itself, which is an open standard, but its EOS operating system and the operational consistency it provides across switching and routing hardware; a customer moving away from Nvidia compute is not necessarily moving away from Arista networking. The open question is whether Anthropic will be the first of many new large buyers, or simply the third name on another concentrated customer list.

In the foundry layer, TSMC receives low risk and is presented as the clearest example that customer concentration and customer power are not the same thing. Its two largest customers accounted for 36% of its revenue in 2025, but the custom silicon efforts that threaten Nvidia's share generate business for TSMC all the same: Google's TPUs, Amazon's Trainium and Nvidia's GPUs compete with each other for compute budgets, but all depend on the same foundry. Whereas fragmentation suits Nvidia so that no handful of customers controls its future, TSMC has the advantage of being able to manufacture both for the incumbent and for the companies trying to replace it; its moat is manufacturing “process power” at extraordinary scale, something neither Samsung nor Intel has yet matched in leading-edge manufacturing and packaging capabilities combined. Nvidia alone accounts for around 60% of advanced CoWoS packaging capacity, and its three largest customers exceed 85% (these are capacity estimates, not shares of TSMC's total revenue). If spending shifts from Nvidia GPUs toward custom accelerators also manufactured at TSMC, the foundry keeps the business; but if total AI infrastructure spending falls, there is no equivalent protection. In short, TSMC is better protected against a change in who wins the AI market than against a contraction of the market itself.

Global Foundries receives medium risk: its two largest customers account for 30% of its wafer revenue. Unlike TSMC, its case is not leadership at the smallest process node, but specialized manufacturing for optical networking, power management, wireless connectivity and other functions where shrinking the transistor is not the only measure of performance. Whichever accelerator wins, the infrastructure will still need to move data and deliver power, so Global Foundries can participate through silicon photonics and related technologies without having to manufacture the winning GPU or ASIC, and it also has diversification opportunities in automotive, IoT and communications infrastructure.

The article closes with a more general reflection. Customer concentration is normally thought of as the risk that a large customer fails, cuts spending or leaves, but the hyperscalers and the frontier labs are the players that recur throughout the entire AI chain, and much depends on their continuing to invest capital and grow revenue. Ben Thompson's Aggregation Theory would partly explain why the same buyers recur across the supply chain: the platforms that won the customer relationship in the internet era enter this new era with distribution, cash flows and workloads that sustain enormous infrastructure investments. But the suppliers below them are not all interchangeable with one another: AI brings concentrated demand into contact with equally concentrated capabilities. Inevitably, as the large buyers grow, they gain both the resources and the incentive to develop alternatives to their suppliers; a customer can become more valuable to Nvidia at the same time as it becomes more capable of replacing it, so growth in the relationship does not guarantee supplier retention. Exposure to continued growth of the hyperscalers and the labs has uneven consequences depending on the supplier: the custom silicon that challenges Nvidia can generate work for Broadcom and wafer demand for TSMC, so one customer's attempt to reduce a supplier's pricing power becomes another supplier's growth opportunity. This explains why concentration risk varies so much even among companies in the same layer: the important distinction is whether a supplier benefits from its customers' growth or becomes the target of their vertical integration. Even so, the author closes by recalling that every player in the chain, without exception, shares the same underlying exposure: that of an eventual slowdown in aggregate AI demand.