I keep coming back to the same uneasy feeling whenever the conversation turns to artificial intelligence spending. Everyone talks about the scale of the opportunity, the infrastructure build-out, the productivity leap that is supposedly just around the corner. Yet a quiet concentration risk has been sitting in plain sight, and one of the investors who called the last major financial collapse is now putting a spotlight on it.
Why Two Startups Suddenly Matter So Much
Steve Eisman, the portfolio manager whose short position against the housing market became famous years ago, recently laid out a concern that feels both obvious and strangely under-discussed. The futures of several of the largest technology companies, he suggested, have become a concentrated bet on the continued success of just two AI firms. Those two names account for a striking share of AI-related revenue at the biggest cloud providers and, in some cases, a meaningful slice of their overall cloud business.
That dependency is not abstract. When a handful of customers drive a large percentage of high-margin growth, any stumble, pricing pressure, or competitive shift can ripple outward quickly. I’ve found that markets often price in the upside of a new technology long before they fully price the concentration risk that comes with it. This time feels no different.
The Revenue Concentration Problem
According to the numbers Eisman highlighted, the two leading AI startups now represent roughly seventy percent of AI-related revenue for several major cloud platforms. In some cases they also make up twenty-five to thirty-five percent of total cloud revenue. Those are not niche figures. They sit at the center of the growth narrative that has supported elevated valuations across the sector.
Think about what that means in practical terms. The capital expenditure programs at the largest technology firms have been justified by the expectation of sustained, high-margin demand from advanced AI workloads. If a significant portion of that demand is tied to the commercial success of only two organizations, the margin of safety shrinks. One delayed product cycle, one unexpected regulatory hurdle, or one aggressive pricing move from a competitor can change the trajectory faster than most models currently assume.
In my experience watching technology cycles, this kind of concentration rarely stays comfortable for long. It either resolves through broader adoption that dilutes the dependence, or it becomes a point of vulnerability when the competitive landscape shifts. Right now the second path looks increasingly plausible.
The Chinese Open-Weight Challenge
The most immediate threat Eisman flagged is not a sudden collapse of either company. It is the steady advance of cheaper Chinese open-weight models. These systems are already available at significantly lower cost, and early signals suggest they are beginning to capture market share among cost-sensitive users and developers.
Price competition in infrastructure-heavy industries tends to be brutal once it starts. Cloud compute is no exception. If a growing number of customers discover they can achieve acceptable performance with models that cost a fraction of the leading proprietary systems, the pressure on pricing power becomes real. That pressure does not stay confined to the model providers. It flows upstream to the cloud platforms that have built capacity on the assumption of continued high-value workloads.
I’ve watched similar dynamics play out in earlier technology waves. When a lower-cost alternative reaches a quality threshold that is “good enough” for a large segment of the market, the premium pricing that supported earlier investment can erode quickly. The question is not whether that threshold will be reached, but how soon and how widely.
The Achilles heel of this whole story is if something bad happens to those two companies or if the Chinese open-weight models start really taking a lot of market share. You could have a big price war. And then we have a problem.
That framing is blunt, yet it captures the core risk. The entire capital cycle currently rests on the premise that demand will remain both strong and high-margin. A sustained price war would test that premise in real time.
Echoes From Another Famous Bear
Eisman is not the only investor who called the housing crisis and is now expressing caution about the AI narrative. Another well-known figure from that period has taken an even more direct stance, questioning how much of the current demand is truly end-user driven versus circular financing arrangements among the companies themselves. He has also positioned against some of the largest beneficiaries of the spending wave, including the dominant semiconductor supplier and the broader chip sector.
Whether one agrees with the degree of skepticism or not, the appearance of two independent voices with proven track records raising related concerns is worth noting. Markets have a habit of dismissing such warnings until the data becomes undeniable. By then the adjustment is usually sharper than most participants expected.
Perhaps the most interesting aspect is how little the market has so far penalized the concentration risk. Valuations continue to reflect confidence that the current leaders will maintain their edge and that demand will stay robust enough to absorb whatever capacity is being built. That confidence may prove correct. It may also prove temporary.
What A Price War Would Actually Look Like
Imagine a scenario in which open-weight models continue to improve and adopt faster than many forecasts currently assume. Enterprise customers, always sensitive to cost, begin shifting workloads. Startups and developers who previously paid premium rates for access to leading models switch to lower-cost alternatives. The revenue growth that justified aggressive capacity expansion starts to slow.
Cloud providers would face a difficult choice. They could cut prices to defend volume, compressing margins on the very workloads that were supposed to deliver the highest returns. Or they could accept lower utilization rates on expensive new infrastructure. Neither outcome is attractive for shareholders who have priced in continued high growth and expanding margins.
The semiconductor side of the equation would feel the pressure as well. Chip demand that was predicated on sustained hyperscale build-outs could moderate more quickly than capacity plans allow. Inventory adjustments and slower order rates have a way of arriving suddenly once the demand signal softens.
None of this is inevitable. It is, however, a plausible path that receives far less attention than the bullish scenarios. Risk management, in my view, requires giving the uncomfortable path at least as much weight as the comfortable one.
The Broader Capital Cycle Context
Every major technology wave has eventually confronted the question of returns on capital. The internet build-out of the late 1990s, the smartphone and mobile data expansion of the following decade, and the earlier cloud migration all eventually required proof that the money spent would generate adequate returns. AI is no different in principle, even if the absolute dollars involved are larger.
What makes the current moment distinctive is the speed and concentration of the spending. Capacity is being added at a pace that assumes near-perfect visibility into future demand. When that demand is itself concentrated among a small number of customers whose competitive positions are still evolving, the assumption becomes harder to defend.
I’ve found that the most useful questions at this stage are not the ones about ultimate potential. They are the ones about intermediate fragility. How much of the current revenue is truly sticky? How sensitive are the customers to price? How quickly can alternative models close the remaining quality gap? Those questions do not have precise answers yet, but they deserve more attention than they currently receive.
Investor Implications Beyond The Headlines
For anyone holding significant exposure to the companies most closely tied to the AI infrastructure build-out, the concentration risk is no longer theoretical. It is measurable. Portfolio construction that ignores the shared dependency on a small set of AI customers is, in effect, making a leveraged bet on their continued commercial success.
That does not mean the entire thesis is wrong. It does mean the risk-reward calculation has shifted. Higher concentration usually warrants either lower position sizes or explicit hedges. Few portfolios appear to have made that adjustment yet.
One practical approach is to examine the percentage of cloud growth that is truly diversified versus the percentage that remains tied to the leading model providers. Where the latter is high, the margin of safety is lower than headline growth rates suggest. Another is to watch the pricing trends in the open-weight segment more carefully than most currently do. Early signs of accelerating adoption there would be a leading indicator of pressure to come.
- Track the share of AI-related revenue coming from the top two customers at each major cloud platform
- Monitor pricing and adoption rates of open-weight models in enterprise and developer communities
- Reassess capital expenditure guidance for sensitivity to slower high-margin demand
- Consider whether semiconductor demand forecasts adequately account for a possible moderation in hyperscale growth
- Evaluate overall portfolio concentration across companies that share the same key customers
These steps sound straightforward. In practice they require looking past the compelling narrative and focusing on the customer-level data that most presentations prefer to gloss over.
Why The Market Has Been Slow To React
Part of the reason the concentration risk has not yet weighed more heavily on valuations is simple momentum. The narrative of transformative AI capability remains powerful, and the early revenue results have been strong enough to keep most investors focused on the upside. Another part is the genuine difficulty of modeling the competitive dynamics. Open-weight models improve rapidly, yet the performance gap with the leading proprietary systems is still material in many demanding applications.
Markets also tend to underweight tail risks until they begin to materialize. A gradual loss of pricing power is less dramatic than a sudden failure, so it receives less attention in the short term. Yet gradual erosion can still produce significant valuation adjustments once it becomes clear that earlier growth assumptions were too optimistic.
In my view the current calm should not be mistaken for the absence of risk. It is more likely the calm that precedes a period of closer scrutiny. When that scrutiny arrives, the companies with the most diversified customer bases and the most flexible cost structures will be better positioned than those whose growth remains heavily dependent on a narrow set of high-value relationships.
A Longer Historical Lens
Technology history is full of moments when a small number of customers or applications drove an outsized share of industry growth, only for that concentration to become a liability later. Mainframe computing, early internet infrastructure, and certain segments of the smartphone supply chain all experienced versions of this pattern. The companies that survived and thrived were usually the ones that used the high-growth period to broaden their customer base and improve cost efficiency before the inevitable competitive response arrived.
Whether the current AI infrastructure leaders will follow that path remains an open question. The capital intensity of the present build-out leaves less room for error than some earlier cycles. The speed at which open-weight alternatives are improving also compresses the window in which the leading providers can lock in durable advantages.
None of this means the technology itself is overhyped. The capability gains are real and the long-term applications are likely to be substantial. The issue is the translation of those capabilities into sustainable returns on the enormous capital being deployed right now. That translation is far from guaranteed, especially when so much of the near-term revenue rests on so few customers.
Practical Questions For Portfolio Managers
Anyone managing capital through this period should be asking a set of specific questions that go beyond the usual growth forecasts. How much of the expected free cash flow in the outer years depends on the continued commercial success of the two leading model providers? What happens to utilization rates and pricing if open-weight alternatives capture even a moderate share of the addressable market? How flexible are the capital expenditure plans if demand growth slows by several hundred basis points?
These are not bearish questions by default. They are simply the questions that risk-aware investors should be able to answer with some precision. Right now many presentations still treat the demand outlook as relatively inelastic and the competitive landscape as relatively stable. Both assumptions look increasingly open to challenge.
I have also noticed that the discussion of circular revenue arrangements remains underdeveloped in most mainstream analysis. When companies are both suppliers and significant customers of one another, reported growth rates can become less informative about true end-market demand. Sorting the genuine external demand from the internal recycling is difficult from the outside, yet it is essential for anyone trying to assess the durability of the current cycle.
Where The Real Vulnerability Lies
The core vulnerability is not that artificial intelligence will fail to deliver value. It is that the current financial architecture supporting the boom is more fragile than the technology itself. Two companies sit at the center of a large share of the high-margin revenue. Cheaper alternatives are already improving. The capital being spent assumes that pricing power and volume growth will both hold. If either assumption weakens, the adjustment process could be uncomfortable for equity holders who have paid for a more linear outcome.
Eisman’s warning is useful precisely because it is not apocalyptic. It simply identifies a single point of failure that most models treat as robust. Treating it as robust is a choice, not a fact. Choices of that kind have consequences when the data eventually diverge from the assumption.
For now the market continues to favor the optimistic path. That may prove correct for longer than skeptics expect. It may also reverse more abruptly than optimists are prepared for. The difference between those two outcomes will likely turn on how quickly and how broadly the cheaper open-weight models continue to gain traction, and on whether the leading proprietary providers can maintain enough differentiation to justify premium pricing.
Those two variables are worth watching more closely than almost any other near-term indicator in the technology sector. They will tell us whether the current concentration is a temporary feature of an early market or a structural weakness that eventually forces a reset in expectations.
Balancing Optimism With Realism
None of the above should be read as a blanket dismissal of the AI opportunity. The underlying technology is advancing rapidly and the potential applications remain vast. The issue is the financial structure that has grown up around it in a relatively short period. That structure contains a concentration risk that is now large enough to matter for the valuation of some of the world’s most important companies.
Healthy skepticism does not require predicting collapse. It requires acknowledging that the path from today’s spending to tomorrow’s returns is narrower and more dependent on a small number of commercial outcomes than many investors currently assume. Narrow paths can still be navigated successfully. They simply leave less room for error.
In the end the market will decide how much weight to give this particular Achilles heel. Until more data arrives, treating the concentration as a material risk rather than a footnote seems the more prudent course. The investors who recognized similar vulnerabilities in earlier cycles were not always the most popular in the moment. They were often the ones who still had capital left when the cycle turned.
That lesson remains relevant. The technology may be new. The pattern of concentrated risk inside a high-growth narrative is not.
