In partnership with

Spend enough time reading AI headlines and you'd think enterprises are slamming on the brakes.

Stories about developers burning through token budgets, companies imposing monthly spending caps, and employees losing access to premium models have created the impression that enterprise AI spending is running into a wall.

SemiAnalysis recently spoke with more than 50 enterprise customers to understand how organizations are managing AI token spend. What emerged wasn't a picture of companies spending less on AI. It was a picture of companies becoming much more deliberate about where AI creates value. Budgets are becoming part of normal operations, even as AI investment continues to grow.

What's interesting is that there doesn't appear to be a playbook yet. Some organizations cap employees at a few hundred dollars a month. Others allow several thousand depending on role. Some have no formal limits at all. The number itself seems less important than the exercise. Enterprises are trying to understand how much AI different jobs actually need and where additional spending continues to improve business outcomes.

There's another reason those conversations are becoming more important.

The instinct is to assume that as AI models become faster and less expensive, infrastructure costs should become easier to manage. Instead, the opposite is happening. Lower token prices are making AI economical for more teams, more applications, longer context windows, and increasingly autonomous agents. Every improvement in efficiency creates another reason to use AI more often.

Economists have a name for this phenomenon. Jevons paradox describes what happens when efficiency improvements increase total consumption instead of reducing it. WEKA's Val Bercovici has argued that AI is becoming one of the clearest modern examples. Cheaper tokens don't reduce infrastructure demand. They unlock new workloads that were previously too expensive to justify.

That dynamic helps explain why enterprises are spending so much time thinking about governance. The challenge isn't simply controlling token spend. It's deciding which workloads require frontier reasoning models, which can run on smaller or open-weight alternatives, and how to keep increasingly autonomous systems from consuming expensive compute where it adds little value.

The organizations making the largest AI investments tend to have one thing in common. They're spending where AI directly affects output. Engineering teams, data scientists, and technical researchers consistently receive larger budgets because their work benefits from iterative reasoning, code generation, and large-context analysis. Other functions often don't require the same level of compute. Instead of treating AI like another software license, companies are beginning to allocate intelligence much the way they allocate any other strategic resource.

The same thinking is showing up in infrastructure decisions. Many organizations are changing their default models, reserving premium reasoning models for complex work while routing routine tasks to lower-cost alternatives. Others are investing in better caching and memory architectures so they aren't repeatedly paying to generate the same outputs. Those decisions reduce unnecessary spending without reducing AI adoption.

The productivity gains explain why budgets continue to expand despite greater scrutiny. SemiAnalysis highlighted examples ranging from recruiting workflows that were cut dramatically to analytical work that dropped from days to hours. Those kinds of improvements change the economics. Once a workflow demonstrates measurable value, the question shifts from "Should we spend money on AI?" to "Where should we deploy more of it?"

Enterprise AI hasn't reached a spending ceiling. It's reaching a planning phase. Budgets, model routing, memory architecture, and infrastructure strategy are all becoming part of the same conversation because they're responding to the same reality. As AI becomes cheaper, organizations don't consume less intelligence. They find more places to apply it. The companies that recognize that dynamic early will be better positioned to build infrastructure for where enterprise AI is going, not where it has been.

The infrastructure powering that shift takes center stage at AMD Advancing AI.

Join us July 22-23 at Moscone West in San Francisco.
Register today while space is still available.

Bay Area Startups Collectively Secured $2.7B+ in July MTD

July continued its slow start, with just two megadeals, SambaNova's $1B Series F the largest. But the number of deals was half the usual and begs the question, is this the summer slowdown, or the lull before the storm?

AI Revenue Gap: Sequoia's David Cahn was one of the first people to dig into projected AI infrastructure spend and the revenue needed to pay for it. Three years ago - when NVIDIA was reporting $50B in annual GPU revenue - he pegged it at $200B in AI revenue. Today, he's estimating that AI infrastructure spend in 2026 will hit $1.5T - and puts the corresponding AI revenue number at $300 trillion. Although OpenAI and Anthropic are both reporting ARR in the tens of billions, that leaves a big gap in revenues as the hyperscalers continue to build and take on more debt.

For startups raising capital: Don't waste time on lists, find investors in companies like yours with LinkSV. Stay on top of who's raising, who's closing and who's investing with the Pulse of the Valley weekday newsletter. Founders get the newsletter, database and alerts for just $7/month ($50 value). Check it out and sign up here.

Follow LinkSV on LinkedIn to stay on top of SV funding intelligence, new tech and the companies, investors and executives impacting the startup ecosystem.

Early Stage:

• Prime Intellect closed a $130M Series A, makes frontier AI training accessible to every company.

• Bespoke Labs closed a $40M Series A, a research lab building the environments and infrastructure that make AI agents reliable.

• Hostie closed a $12M Series A, the AI-powered Virtual Concierge for restaurants.

• Piq Energy closed a $5M Seed, helps companies connect projects to the grid faster with an agentic grid planning platform.

Growth Stage:

• SambaNova Systems closed a $1000M Series F, developer of the SN50 RDU (Reconfigurable Dataflow Unit), an AI processor purpose-built for agentic AI inference at lower cost, and higher energy efficiency than GPUs.

• Ollama closed a $65M Series B, makes open models easy, efficient, and safe to use on a user's own device or in Ollama’s cloud.

• Super closed a $65M Series D, the all-in-one app to save, earn and put more money in your pocket.

• Hippo Harvest closed a $30M Series C, a leafy greens company growing USDA-certified organic produce in robotics- and machine learning-powered greenhouses.

  

NeoCloud Council Executive Dinner

Some of the most valuable conversations in AI infrastructure never happen on stage.

They happen around a dinner table.

The NeoCloud Council Executive Dinner brings together founders, cloud providers, enterprise leaders, investors, and infrastructure operators for an evening designed around discussion instead of presentations.

Hosted during AMD Advancing AI Week, the dinner is intentionally small. It's creating the kind of room where people can compare notes on what's actually happening across AI infrastructure, from deployment challenges and customer demand to where the market is heading next.

The people building this industry rarely get the chance to talk candidly with peers facing the same decisions. That's what this dinner is for.

If you're helping shape the next generation of AI infrastructure, this is a conversation worth being part of.

Your Feedback Matters!

Your feedback is crucial in helping us refine our content and maintain the newsletter's value for you and your fellow readers. We welcome your suggestions on how we can improve our offering. [email protected] 

Logan Lemery
Head of Content // Team Ignite

Speak naturally. Send without fixing.

Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.

Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.

89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.

Start flowing free