Date: 2026_05_24 Source: https://www.youtube.com/watch?v=Poyi6X7rOwY Duration: 1417 Platform: YouTube Creator: AI News & Strategy Daily | Nate B Jones

Why the AI boom is about to hit a wall¶

Executive Summary¶

On Microsoft's Q3 earnings call, Satya Nadella told investors the company will spend $190 billion in capex this year and still expects to be capacity constrained through year end. The most valuable software company on the planet — with $190 billion to spend — cannot get enough capacity to meet its own demand. The supply problem is not just GPUs; it's the entire layer below the GPU: whether you can manufacture enough chips packaged with the memory they need to keep up with modern AI workloads. The conclusion: your AI vendor contract is effectively a supply contract in everything but name, and most software companies are not ready for this conversation.

The Core Thesis: AI Is a Factory, Not Software¶

Every answer from a model is the output of a production chip system. The visible product may look like software — ChatGPT, Copilot, Gemini, Claude — but the constraint underneath goes all the way down to the metal. Chips run the math. High bandwidth memory feeds the chips. Packaging connects them together. Networking moves data across the cluster. Power keeps the racks alive. Cooling keeps them at temperature. Operations keeps the whole thing utilized. A user sees a paragraph generated on a screen, but every word in that paragraph came out of a factory.

The hyperscalers have figured this out and are spending accordingly: - Microsoft: $190B capex, still capacity constrained - Meta: $125–145B (raised guidance because component prices are up and they need more data centers) - Amazon: Landed more than 2.1 million AI chips in the last 12 months; multi-gigawatt commitments from Anthropic and OpenAI; more than 1 million Nvidia GPUs deploying through 2027 - Google: $185B in spend last year

The pattern is bigger than any single company. We need to stop treating these companies as software companies and start treating them as physical infrastructure companies. That is how their unit economics work now.

The AI Factory Stack (Layer by Layer)¶

1. Chips¶

The unit of infrastructure is not a GPU or TPU — it's the module. Nvidia's GB200 NVL72 is the canonical example: a liquid-cooled rack-scale system connecting 72 Blackwell GPUs and 36 Grace CPUs into a single NVLink domain, with 13.5 terabytes of HBM3 memory and 576 TB/s of memory bandwidth. A chip alone doesn't produce intelligence at scale — it needs memory close to it, packaging, networking, and a place to run.

2. Memory (The Bottleneck)¶

High Bandwidth Memory (HBM) is the single most constrained input in the entire supply chain. If you cannot move data fast enough, all your compute sits idle. A company can have plenty of GPUs on paper and still not be able to ship usable AI accelerators because they cannot get enough HBM. The four largest AI chip designers consumed: - 90% of global chip packaging capacity - 90% of HBM memory supply - But only 12% of advanced logic die production

In other words: the bottleneck was never the ability to design better chips. It's the ability to turn all of this into an integrated compute supply that enables real tokens to be served at scale.

3. Packaging¶

You have to integrate logic dies and HBM stacks into a single working chip package. TSMC's CoWoS is what connects compute and memory at the bandwidth AI workloads need. Underneath the packaging are substrates and interposers — pieces that carry signals and hold components in alignment. If substrate yield drops, the production line slows down even if the chip design is excellent.

4. Networking / Optics¶

Large AI clusters are communication machines as much as compute machines. The GPUs need to move enormous amounts of data back and forth between one another. Copper has limits at scale — around heat, distance, and signal integrity. At hundreds of thousands of GPUs, the network has to be optical. Nvidia's SpectrumX Photonix announcement represents this shift toward shipped product-scale optical networking.

5. Power¶

The real constraint is firm power at the right location on the right schedule. The country may have plenty of power on paper, but a specific site may not get the power it needs to stand up a data center in time. The IEA projects global data center electricity consumption will roughly double to about945 TWh by 2030, but the real issue is local, site-specific power availability. Even transmission and interconnection for power can stretch construction schedules past 18 months into the four-year range.

6. Cooling¶

Dense AI racks generate heat at levels old data center designs were not built to handle. Liquid cooling is part of production capacity today. If cooling cannot handle the rack density, the hardware doesn't run at full power.

7. Construction / Land / Data Centers¶

Traditional 12–18 month data center timelines are no longer useful for500+ megawatt AI campuses. Meta's Hyperion campus in Louisiana — a joint venture with Blue Owl Capital — is already a multi-year construction project. Every one of these layers has different supply chain players and different timelines. Any one of them can be the bottleneck that determines whether your AI strategy delivers.

The AI Vendor Contract Is a Supply Contract¶

Six months ago, an AI vendor contract was structured like a software contract. Now, with hyperscalers spending at this scale and still rationing heavily, your AI vendor contract is effectively tied into the hyperscalers. It has allocation. It should have capacity terms. It should have fallback. It should have line items that didn't exist a few months ago.

The presenter argues: - Developers are underrepresented in AI procurement decisions - Engineers need to be at the table because they can speak to whether what is being allocated is actually usable - If you don't get these terms right, you can't roll AI out to developers correctly, can't roll it out in AI operations correctly, and you will run out of capacity when you really need it

Key Strategic Implications¶

1. Stop Thinking of AI as Software with a Fancy Backend¶

The visible product may look like software, but the constraint underneath is all the way down to the metal. Every answer from a model is the output of a production chip system.

2. Your AI Vendor Contract Is a Supply Contract¶

AI vendor agreements now sit on top of a physical supply chain. Buyers feel embarrassed to ask about allocation, capacity, delivery, and fallback — vendors don't want to talk about it because they may not have full answers. You need to be in a position to have honest conversations and acknowledge the uncertainty.

3. The Bottleneck Is Below the Chip Level¶

It is not design bandwidth (12% of the world's advanced logic die production supports 90% utilization of packaging and memory). The bottleneck is the ability to turn chips, memory, packaging, networking, power, and cooling into an integrated compute supply that serves real tokens at scale.

4. This Is an Infrastructure Shift, Not a Software Shift¶

The hyperscalers' decision to move heavily into physical infrastructure is shaping the intelligence the rest of us get — either directly from them or from vendors who build on top of their stacks. When a vendor says they're "investing in AI infra," they typically mean the thin layer on top of this whole factory system.

5. The Useful Executive Question¶

Not "who benefits from AI capex" (that becomes stock picking fast), but: "Where in the supply chain does a delay stop you from shipping AI?" This is the question that should drive procurement and contract strategy.

🦐 Summary by Thrawn the Prawn — Strategic Analysis Division