The Market Is Shifting From "Who Spends" to "Who Controls Scarcity"
The market's thinking about artificial intelligence has matured. In the earliest stage of the cycle, the obvious and easiest money came simply from identifying who was spending the most on AI and buying the most apparent bottlenecks early. That phase is giving way to something more discriminating. The market is now beginning to reward the companies that actually control the scarce inputs — the contract capacity, the data movement, the power infrastructure, the edge — all the layers that make the entire AI economy function.
Eventually, this dynamic is expected to return toward the normality of the next phase of the AI cycle, where the market will start separating genuine, durable franchise value from temporary scarcity. Until that normalization fully arrives, the non-"Magnificent 7" AI winners are likely to remain a specific set of bottleneck plays.
The Five Bottlenecks of the Non-Mag 7 AI Basket
The core thesis centers on what can be called the "complete five bottleneck" non-Mag 7 AI basket — five layers where supply is structurally constrained and where investment opportunity remains:
1. Memory — The most obvious of the bottlenecks, underscored most recently by Micron's earnings. Memory is the clearest example of demand running far ahead of supply.
2. Connectivity — A critically important bottleneck. Data movement constraints inside AI factories (data centers) will only continue to grow in importance as workloads scale.
3. Power — Set to become a major physical constraint on how fast AI data centers can actually be built and scaled. Electricity availability is a hard ceiling on growth.
4. Compute capacity — Functioning as the overflow layer for when hyperscalers are sold out. Because hyperscaler capacity is fully booked, this overflow demand is expected to persist, and it feels like it will be years before supply and demand even out.
5. On-device CPU — The most recent bottleneck to emerge, described as the next "leg" of AI's graduation. The industry is moving from cloud-only training toward inference running on agents, PCs, and phones. This inference breakout is creating a new CPU bottleneck, which is why Intel, AMD, and ARM have all performed exceptionally well recently.
Are These Bottlenecks Investable, and for How Long?
A clarifying question was raised: are these five groups — memory, connectivity, power, compute, and CPUs — liked as sectors, or are they names to watch?
The answer was that they are genuinely liked, not merely watched. The bottlenecks are expected to last for years. Assigning a "peak earnings" to any of these bottlenecks has been premature. Every three months the market is "soberly reminded" that the bottleneck continues and that supply is nowhere near demand. Micron's own commentary is the evidence: the company has no visibility into when supply and demand will line up. Customers are now signing strategic five-year contracts that guarantee a minimum pricing floor. This contractual behavior signals that there are years of bottleneck conditions ahead, and that the situation is nowhere close to resolving.
The Losers: Enterprise Software Faces the "SaaS Apocalypse"
While the AI super-cycle has driven enormous investment, a number of headline enterprise software companies are fizzling. The common thread is that AI capabilities are now challenging the core revenue models of established software giants — a phenomenon summarized succinctly as the "SaaS apocalypse."
- Adobe — Disappointing because its main portfolio capabilities, including text-to-video and text-to-image generation, can be cannibalized and undercut by open-source generative AI and other related AI tools. As a result, its Creative Cloud subscriber base is under siege.
- Salesforce — Penalized in valuation since the beginning of the year over concerns that its next-generation autonomous AI agents may not protect the number of seat licenses its large enterprise customers have historically required. If AI agents reduce the need for human seats, the traditional recurring-revenue model is disrupted. Management's narrative has promised double-digit growth, acceleration, and that "Agentforce is doing really well" — but the company is actually delivering only high single-digit organic growth, with no credible reacceleration story in sight.
- Atlassian — Penalized over concerns about its long-term free cash flow, again tied to generative AI coding tools reducing the developer headcounts that underpin its historical revenue success.
- ServiceNow — Impacted for very similar reasons. It had no room for error on margins, and it was forced to announce delays in some notable Middle East contracts, which further hurt performance. These pressures are expected to continue through the rest of the year.
What Defines an AI Loser?
The unifying principle for the losers: software companies that sell features and sell platforms are in danger if they cannot prove they are a "top of funnel" — a destination that new clients or enterprises actively migrate to. If a software company is showing no reacceleration from AI at this stage of the buildout, it is in trouble, and it may be more of an "AI wrapper" than a genuine outcome provider for clients.
The dislocation between winners and losers is widening. On the losing side sit names like Adobe, Monday, Salesforce, and a wide range of different DevOps tools. The current environment is "put up or don't be part of the winning cycle" — companies have to deliver the actual, proven KPIs to be considered winners.
The Winners: Accelerators, Connectivity, and On-Device Silicon
On the hardware and infrastructure side, several names stand out as clear winners:
- Nvidia — Up 26% on the year, remaining a central winner.
- Broadcom and Marvell — At the apex of the AI accelerator space while also streamlining efficiencies inside the data center.
- Qualcomm — Highlighted as "firing on all cylinders" following its recent event in New York City. It is poised to enter the data center market with an impressive portfolio, leveraging its efficiency advantages and its on-device AI design capabilities.
On the software side, the winners are the companies actually proving their value through KPIs — names such as Snowflake, Palantir, and Datadog.
A Note on the Bottleneck Basket's Own Weak Link
Within the favored bottleneck basket — which spans names from Micron to GEV, CoreWeave, and IREN — the candidate for relative weakness, if one had to be chosen, would be IREN.
The Bottom Line
The AI trade is bifurcating sharply. The durable winners are those controlling genuinely scarce physical and computational inputs — memory, connectivity, power, overflow compute, and on-device CPUs — where multi-year contracts and persistent supply shortages point to years of pricing power ahead. The losers are the legacy software platforms whose feature-and-seat business models are being eroded from underneath by the very generative AI wave that was supposed to lift them. In this next phase, scarcity and proven outcomes are rewarded; everything else risks being exposed as a wrapper around someone else's capability.
