JPMorgan Just Challenged the Cloud-Only AI Thesis
Listen to the audio version of this article (generated by AI).
The AI infrastructure story that has dominated markets for the past two years has had one assumed ending: eventually, every enterprise will migrate its AI workloads to the hyperscaler cloud. AWS, Azure, Google Cloud, Oracle (ORCL) — pick your platform, pay per token, and let someone else worry about the hardware.
JPMorgan Chase (JPM) just added an important asterisk.
This week, SambaNova Systems — an AI chip company that Intel (INTC) reportedly tried to acquire for about $1.6 billion less than a year ago — raised $1 billion at an $11 billion valuation.
The customer that made the announcement so interesting was JPMorgan Chase, which selected SambaNova as its inference-infrastructure partner, deploying its systems to power secure, on-premises AI inference at the bank.
The speed of the startup’s re-rating — and who signed on as the anchor customer — isn’t an accident.
JPMorgan Just Put an Asterisk on the Cloud-Only AI Thesis
The mainstream AI infrastructure thesis assumes that inference demand — the workload created every time an AI model answers a query, writes code, or completes a task — primarily flows through hyperscaler cloud platforms.
That’s been true so far, and it will remain true for most of the market.
But JPMorgan’s decision points to a segment that the cloud-first narrative underweights: enterprises and institutions that simply cannot send their most sensitive data to a third-party server.
Banks hold client data and proprietary trading strategies they can’t expose. Hospitals manage patient records that federal law requires them to protect. Defense contractors and government agencies often face outright restrictions on running sensitive workloads on commercial cloud infrastructure.
For these organizations, cloud economics are appealing on paper. But that architecture comes with a data exposure risk they can’t accept.
SambaNova’s CEO framed the JPMorgan win as a signal to the whole banking industry: banks want control over their most sensitive inference, and they’re starting to build for it. And the vendors that give them that control are about to have a very interesting few years.
Why Enterprise AI Inference Looks Different From Chatbots
We’ve written at length about the inference supercycle — the shift from AI as a training-era story to AI as a persistent, always-on workload running inside enterprise operations. Agentic AI is accelerating that shift, with agent-based workflows consuming more compute than single-shot queries ever did.
What SambaNova’s round shows is that the inference supercycle has a niche the market hasn’t fully accounted for.
A meaningful slice of enterprise inference demand won’t flow through hyperscaler APIs. It will run on-premises, inside the firewall, on hardware owned and operated by the enterprise itself.
Liang noted that enterprises and governments are just starting their AI journey, with most growth so far concentrated among tech’s model makers and frontier labs — leaving substantial revenue still on the table. In regulated industries specifically, that revenue goes to whoever sells the hardware, the networking, the storage, and the software stack that makes on-premises inference work.
But the next phase of the AI trade has more moving parts than most investors realize. If you want to hear where I think the smartest money in AI is moving next — my highest-conviction ideas, live and in-person — I’ll be at the Stansberry Conference & Alliance Meeting in Las Vegas later this year. Interested? Reserve your discounted seat before they sell out.
The AI Infrastructure Trade Is Splitting Between Cloud and On-Prem
The picks-and-shovels thesis for AI infrastructure remains intact. The global AI inference market is valued at roughly $120 billion in 2026 and projected to reach $300-plus billion by 2034. That demand has to live somewhere.
Now that “somewhere” is looking a bit more bifurcated.
Hyperscaler cloud captures the majority of it. Within regulated industries, on-premises inference is forming as its own distinct market. Banks, hospital systems, and government agencies can build a compelling economic case for owning their own hardware. The cost per token math favors on-premises at sufficient utilization. And when the regulatory constraints are real, the economics almost don’t matter. Cloud simply isn’t a viable option for their most sensitive workloads.
The names positioned for this are the same ones we’ve been writing about. Dell‘s (DELL) AI Factory already has more than 4,000 enterprise customers. Everpure (P) — formerly Pure Storage — has rebuilt its platform specifically to make enterprise data accessible to AI workloads without the overhead of replication.
JPMorgan’s decision just made their pitch to the next bank a lot easier.
The Bottom Line
SambaNova going from a rumored $1.6 billion acquisition target to raising at $11 billion in under a year reflects something real: private capital has decided that secure, on-premises enterprise AI inference is a durable market, and the price of getting in has changed accordingly.
The frontier labs and hyperscalers drove the first phase of this trade. The enterprise and sovereign deployment wave is the second phase — and within regulated industries, it plays by different rules. Banks, hospital systems, and government agencies don’t move fast. But when they do, they move at scale, under long-term contracts, with infrastructure budgets that tend to be sticky.
Other banks are likely watching JPMorgan’s move. So are certain corners of healthcare and government. For data-sensitive organizations, this could be the new blueprint.
The inference supercycle is real, and the hyperscaler cloud will capture most of it. But within sensitive sectors, a structurally distinct market is forming for secure, on-premises inference infrastructure. For the companies best positioned to serve it, it’s a durable one.
And durable infrastructure spend is exactly what the most sophisticated private capital has been positioning around… not at the application layer or the model layer, but underneath all of it.
The energy systems, nuclear capacity, and physical fabrication that make persistent AI compute possible — whether it runs in a hyperscaler’s data center or inside JPMorgan’s firewall — are being secured through private funds and bilateral agreements that most investors never see.
And though most of those positions aren’t accessible publicly, there are seven publicly traded stocks that mirror those same bets almost exactly — the hard-asset backbone of an infrastructure build that isn’t slowing down regardless of where enterprises decide to run their workloads.