The model layer is a liquidity play dressed up as a growth story—the real alpha is in the picks and shovels that haven't IPO'd yet.
The Summary
- Athena Capital's Isabelle Freidheim says the AI model layer is already built, funded, and repriced in private markets—current public trading is mostly exits, not expansion capital
- Infrastructure deployment is where the growth lives now: data centers, edge compute, networking, power delivery
- Public market investors should wait for infrastructure names to IPO over the next few years for earlier-stage exposure to AI buildout
The Signal
The AI investment thesis just split in two. On one side: the model makers, already fat with capital and trading at fully discovered prices. On the other: the infrastructure layer, still private, still building, still offering actual growth potential instead of just liquidity.
Freidheim's framing cuts through the confusion around AI public offerings. When a foundation model company goes public in 2026, you're not buying the future. You're buying someone else's past. The valuation happened three funding rounds ago. The risk got priced in Series C. What looks like a hot IPO is really just early investors finding an exit while retail buys the story.
"Current public market activity around AI models largely represents liquidity events rather than new capital raising."
This matters because infrastructure scales differently than models. A foundation model gets trained once, then serves inference at decreasing marginal cost. Infrastructure? That's linear buildout. Every new model, every million new agents, every enterprise deployment needs physical compute, networking, power delivery, and cooling. The model layer concentrates. The infrastructure layer sprawls.
The timing insight is the real signal here. Infrastructure companies are still private because they're still building. Data center developers, chip packaging specialists, power grid modernizers, edge computing networks—these businesses need years and billions before they're IPO-ready. Which means public market investors haven't had clean access yet.
Key infrastructure plays coming to market:
- Regional AI data center operators building closer to enterprise customers
- Power delivery and grid infrastructure serving compute clusters
- Networking companies handling east-west traffic between AI workloads
- Cooling and thermal management for dense GPU deployments
Compare this to buying Nvidia in 2016 versus buying it in 2024. Same company, radically different risk/reward profiles. Early infrastructure exposure means you're buying capacity before demand fully materializes. Late model exposure means you're buying hype after product-market fit is already priced in.
The agent economy accelerates this dynamic. Every company spinning up AI agents needs somewhere to run them. Every developer building autonomous systems needs reliable compute that scales. The infrastructure layer captures that growth directly. The model layer? They compete on price and capability while infrastructure providers just collect rent on the buildout.
The Implication
If you're deploying capital into AI in public markets, the message is clear: wait for the infrastructure wave. The model layer already ran. Track which private infrastructure companies are approaching IPO scale. Watch for data center operators, power delivery specialists, and edge compute networks filing S-1s over the next 18-24 months.
For builders in the agent space, this validates the thesis that compute availability matters more than model sophistication. The bottleneck isn't model capability anymore. It's deployment infrastructure. If you're running agents at scale, your real competitive advantage is access to reliable, low-latency compute infrastructure. Lock in those partnerships now while the public markets are still pricing infrastructure at private multiples.