AI’s Real Advantage Lies in Foundational Layers

Alex Sacerdote of Whale Rock Capital sees the AI revolution not as a single wave but as a layered system of interlocking s-curves--each with its own timing, winners, and hidden consequences. His conviction in Anthropic isn’t just a bet on a model, but on an entire stack that’s redefining enterprise productivity, developer output, and infrastructure demand. The non-obvious implication? The biggest gains aren’t in the most visible layer (applications), but in the foundational layers where feedback loops compound: chips, infrastructure, and models that improve themselves. This post reveals how Sacerdote maps these dynamics, why most investors miss them, and where the real advantage lies--not in predicting the future, but in understanding how systems evolve. If you're an investor, operator, or strategist trying to separate signal from hype in AI, this analysis exposes the structural shifts that will outlast headlines.

Why the Obvious Fix--Building AI Apps--Creates a Hidden Bottleneck

Most investors and entrepreneurs look at AI and see an opportunity to build new applications: smarter CRMs, next-gen productivity tools, AI-native workflows. But Alex Sacerdote sees a different reality. The immediate action--throwing engineers at AI apps--is creating a downstream bottleneck: compute scarcity. Every new AI agent, every coding assistant, every enterprise workflow built on top of foundational models consumes tokens, which require GPUs, which require power, cooling, networking, and memory. And right now, there isn’t enough.

"There's not going to be enough compute. I'm so curious when an investor like you who historically was a public markets investor you could hit buy and buy whatever you want is now operating in lots of the most important private market companies."

-- Patrick O'Shaughnessy

Sacerdote’s shift from public to private markets wasn’t just opportunistic--it was systemic. He realized that the companies controlling the means of production in AI aren’t the ones building end-user apps, but the ones building the underlying stack. When Anthropic’s developers were spending $100 a day on tokens, that wasn’t just a cost--it was a leading indicator of demand. Multiply that by millions of developers, and you get a half-trillion-dollar coding market alone, running on infrastructure that’s already strained.

This creates a feedback loop: more AI usage → more demand for compute → more pressure on hardware → higher prices and longer lead times → only the best-capitalized players can scale. The immediate benefit of building an AI app is obvious. The hidden cost is that most apps can’t scale because the infrastructure can’t keep up. And that’s where the competitive advantage shifts--not to the fastest mover, but to the deepest-pocketed, best-connected, and most strategically embedded.

The system responds by rewarding companies that are already in the foundation: Nvidia, TSMC, ASML, Celestica. These aren’t just beneficiaries of demand--they’re gatekeepers. And because their lead times are years, their moats are widening just as the surface-level AI frenzy peaks.

The De-Commoditization of Hardware: When “Good Enough” Is No Longer Good Enough

For decades, the data center was a commodity business. Intel x86 chips, standard memory, off-the-shelf networking--everything improved slowly, predictably. Moore’s Law delivered incremental gains, and innovation stalled. But AI has shattered that equilibrium.

Workloads are growing 10x per year, not 25--40%. That kind of demand doesn’t just strain systems--it forces reinvention at every layer. Memory isn’t just memory anymore. High Bandwidth Memory (HBM) stacks 10 chips vertically. Networking isn’t just 100G to 400G over seven years--it’s upgrading annually. Power supplies need to handle 50--125x more load. Cooling isn’t air--it’s liquid. And printed circuit boards? They’ve gone from 10 layers to 40, with only a handful of suppliers capable of producing them.

This isn’t just growth. It’s de-commoditization.

"We call it the de-commoditization of the hardware industry... all the companies are public and they all have powerful IP."

-- Alex Sacerdote

What makes this shift so hard for conventional investors to grasp is that it contradicts decades of tech history. Since the cloud commoditized servers, the assumption has been that hardware is a low-margin, undifferentiated business. But AI breaks that model. When performance bottlenecks are everywhere, differentiation isn’t optional--it’s existential.

Take Celestica. Once a struggling contract manufacturer, it’s now the sole supplier of Google’s TPU servers. Why? Because it retained its supercomputing heritage from IBM and mastered liquid cooling--a capability most competitors failed to replicate. Now, its servers cost $200K--$300K each, and if one fails, the entire cluster goes down. That’s not a commodity. That’s a mission-critical component with no easy swap-in.

The consequence? Companies like Celestica, Corning, and Delta Electronics are no longer price-takers. They’re partners in roadmap design, with visibility into demand four years out. Their margins are rising. Their growth is accelerating. And their valuations are still catching up to the reality that they’re no longer vendors--they’re enablers.

The kicker? This isn’t a temporary cycle. It’s structural. Even if AI adoption slowed, the hardware wouldn’t revert to commoditization. The genie is out of the bottle: enterprises now know what’s possible, and they won’t go back to “good enough.”

The Recursive Advantage: When Models Improve Themselves

Most AI analysis stops at demand: more users, more tokens, more revenue. But Sacerdote sees a deeper dynamic--one where the leading models don’t just benefit from scale, they become better because of it.

Anthropic’s lead in coding isn’t just a market position. It’s a feedback loop. Every time a developer uses Claude Code, they generate data--code outputs, corrections, usage patterns. That data is fed back into the model, improving its performance, especially in enterprise-critical tasks. Over time, this creates a recursive advantage: better model → more usage → more data → better model.

This is not a race to zero. It’s a race to escape velocity.

"If you're 80% close to the top of the benchmarks, going from 80 to 85 is a huge unlock. The open source guys--they don't have as much compute, so they can come close to the leading edge but they can't leapfrog it."

-- Alex Sacerdote

Open source models can mimic, but they can’t replicate the closed-loop improvement of private, well-funded labs. They lack the compute, the data, and the capital to sustain the pace. And because scaling laws still hold, the gap isn’t narrowing--it’s widening.

This changes the competitive landscape. It’s not just about who has the best model today. It’s about who can fund the next three generations of training runs, who can attract the best talent, and who can build the ecosystem around the model.

Anthropic isn’t just selling an API. It’s building a “harness” of tools--SDKs, orchestration layers, enterprise integrations--that deepen lock-in. Just like AWS didn’t win by selling servers, but by inventing services no one else saw coming, Anthropic is playing the same game: turn a model into a platform.

The delayed payoff? Most investors still see AI models as commodities. They’ll be shocked when they realize the leaders aren’t interchangeable--they’re irreplaceable.

The L-Curve of Enterprise Adoption: Why It Feels Slow Now But Will Flip Fast

One of the most dangerous misconceptions in tech investing is assuming adoption is linear. Sacerdote sees AI not on an S-curve, but on an “L-curve”--flat for a long time, then a near-vertical takeoff.

Today, Sundar Pichai says AI is at “10 basis points” of knowledge workers. Anthropic has 14--15 million DAUs, but only a fraction are using it in agentic ways. Most enterprises are still in “internet 1.0” mode--knowing they need AI, but not knowing how to build it.

But that’s about to change.

The trigger? Tools like Claude Code have crossed a threshold. Andrej Karpathy and Linus Torvalds--two of the most respected coders in the world--have stopped writing code in traditional languages. They’re using AI to express intent in English. That’s not incremental. That’s transformative.

And because knowledge work is highly networked, adoption will cascade. Once a few teams in a company go all-in on AI agents, the rest will follow--fast. The cost of not adopting will be obsolescence.

This is where the system routes around resistance. CIOs worried about security or job displacement will be overruled by business units desperate for productivity. The same way the CIA adopting AWS broke the cloud security taboo, early AI wins in engineering and finance will break the enterprise inertia.

The consequence? The market isn’t 10% adopted. It’s 0.1% adopted in meaningful ways. And when the flip happens, it won’t be gradual--it’ll be a light switch.

Key Action Items

• Over the next quarter: Audit your AI infrastructure exposure. If you’re not directly invested in chipmakers (Nvidia, TSMC), memory (SK Hynix, Micron), or advanced packaging (ASE, Amkor), you’re missing the foundation. Start building positions in companies with multi-year design wins.

• Within 6 months: Re-evaluate enterprise software incumbents using the “Modified Rule of 40”: (AI % of revenue) + (market share in AI category). Companies below 20 on this score are at risk of displacement. Prioritize those actively integrating AI into core workflows, not just bolt-ons.

• This pays off in 12--18 months: Build relationships with private AI infrastructure players. The next wave of value creation will be in edge cases--liquid cooling, optical interconnects, AI-optimized networking. These aren’t public yet, but they will be.

• Start now, discomfort required: Shift analyst time from app-layer hype to deep technical due diligence on hardware and model layers. This means learning about HBM, CXL, photonic interconnects--topics most investors avoid. The advantage comes from going where others won’t.

• Long-term (2+ years): Bet on recursive improvement. Invest in companies where usage directly improves the product--Anthropic, OpenAI, maybe Google. Avoid pure API plays without ecosystem lock-in or data feedback.

• Immediate: Stop thinking of AI models as commodities. The differentiation in training methods, enterprise focus, and private IP is real. Use management quality--turnover, focus, execution--as a leading indicator.

• Ongoing: Maintain a “learning machine” approach. Sacerdote’s edge isn’t one insight--it’s 20 years of compounding knowledge across cycles. Institutionalize deep dives, management access, and cross-cycle pattern recognition. That’s the real moat.