Listen to the audio version of this article (generated by artificial intelligence).
The AI infrastructure story that has dominated markets over the past couple of years had an assumed ending: Eventually, every organization will migrate its AI workloads to the hyperscale cloud. AWS, Azure, Google Cloud, oracle (ORCL) – Choose your platform, pay per token, and let someone else worry about the hardware.
JPMorgan Chase (JBM) I just added an important asterisk.
this week, Sambanova Systems – A company working in the field of artificial intelligence chips Intel (Intech) It reportedly attempted to acquire about $1.6 billion less than last year – it raised $1 billion at an $11 billion valuation.
The client that made this announcement interesting is JPMorgan Chase, which chose SambaNova as its inference infrastructure partner, deploying its systems to run secure inference and AI within the bank.
The speed with which the startup has been rebranded — and who has signed on as a major client — is no coincidence.
JPMorgan Just put an asterisk on the cloud AI thesis
The prevailing infrastructure premise for AI assumes that inference demand — the workload generated every time an AI model answers a query, writes code, or completes a task — flows primarily through hyper-scalable cloud platforms.
This has been true so far, and will continue to be true for most markets.
But JPMorgan’s decision points to a segment that is underestimating the cloud-first narrative: companies and organizations that simply cannot send their most sensitive data to a third-party server.
Banks keep customer data and trading strategies private that they cannot disclose. Hospitals manage patient records that federal law requires them to protect. Defense contractors and government agencies often face explicit restrictions on running sensitive workloads on commercial cloud infrastructure.
For these organizations, the cloud economy is attractive on paper. But this structure comes with the risk of exposure to data they cannot accept.
SambaNova’s CEO framed JPMorgan’s win as a signal to the entire banking industry: Banks want to take control of their most sensitive conclusions, and they’re starting to build toward it. And the vendors who give them this control are about to have a very interesting few years.
Why enterprise AI inference looks different from chatbots
We’ve written at length about Super cycle inference – Shifting from AI as a training-era story to AI as an ongoing, ongoing workload within enterprise operations. Agentic AI is accelerating this transformation, as agent-based workflows consume more compute than individual queries.
What the SambaNova round shows is that the heuristics supercycle has a place that the market has not fully taken into account.
No meaningful slice of enterprise inference demand will flow via hyperscale APIs. It will run locally, inside the firewall, on devices owned and managed by the organization itself.
Liang noted that companies and governments are just beginning their AI journey, with most of the growth so far concentrated among technology model makers and leading laboratories — leaving significant revenue still on the table. In regulated industries specifically, that revenue goes to whoever sells the hardware, networking, storage, and software suite that makes internal reasoning work.
But the next phase of AI trading has more poignant parts than most investors realize. If you want to hear where I think the smartest money in AI will go next — my most convinced thoughts, first-hand and in person — I’ll be at the Stansbury Alliance Conference and Meeting in Las Vegas later this year. interested? Reserve your discounted seat before tickets run out.
The AI infrastructure trade is divided between cloud and on-premises
The pick-and-pick theory of AI infrastructure remains sound. The global AI inference market is estimated to be worth $120 billion in 2026, and is expected to reach more than $300 billion by 2034. This demand has to go somewhere.
Now “somewhere” seems more complex.
Hyperscaler cloud takes up most of it. In regulated industries, internal reasoning is shaping up as its own distinct market. Banks, hospital systems and government agencies can build a compelling economic case for owning their own devices. Cost per mathematical symbol is preferred locally when sufficiently used. When regulatory constraints are real, economic factors are almost irrelevant. The cloud is simply not a viable option for more sensitive workloads.
The names used for this are the same names we wrote about. Dale(Dale) AI Factory already has more than 4,000 enterprise customers. Everbury (p) — formerly known as Pure Storage — has rebuilt its platform specifically to make enterprise data accessible to AI workloads without the burden of replication.
JPMorgan’s decision made their pitch to the next bank much easier.
Bottom line
SambaNova’s move from a rumored $1.6 billion acquisition target to raising $11 billion in less than a year reflects something real: private capital has decided that secure inference and artificial intelligence for local enterprises is a permanent market, and the price of entry has changed accordingly.
Frontier laboratories and high-tech specialists led the first phase of this trade. The institutional and sovereign diffusion wave is the second phase – and within regulated industries, they play by different rules. Banks, hospital systems and government agencies are not moving quickly. But when they do, they move at scale, under long-term contracts, with infrastructure budgets that tend to be fixed.
Other banks are likely to watch JPMorgan’s move. So do some aspects of health care and government. For data-sensitive organizations, this may be the new blueprint.
The inference supercycle is real, and the hyperscale cloud will catch most of it. But in sensitive sectors, a structurally distinct market for secure inference infrastructure is forming. For companies better positioned for its service, it’s a solid company.
And permanent infrastructure spending is exactly what the most sophisticated private capital has focused on… not at the application level or the model layer, but Underneath it all.
The energy systems, nuclear capacity, and physical manufacturing that make continuous AI computing possible — whether running in a hyperscaled data center or inside JPMorgan’s firewall — are secured through private funds and bilateral agreements that most investors never see.
And while most of these positions aren’t publicly accessible, there are seven publicly traded stocks that reflect roughly the same bets — a hard asset backbone for building infrastructure that doesn’t slow down no matter where companies decide to run their workloads.
