GM. This is Milk Road AI, the newsletter that spots the bottleneck before the rest of the market even knows traffic exists.

​Here’s what we’ve got for you today:

Consensus Miami is one of the largest digital asset conferences that’s going all in on crypto and agentic commerce. Grab your passes at 20% off.

THE GREAT CPU FAMINE

Napoleon Bonaparte had the most brilliant military mind in European history.

He had superior tactics, stronger generals, and elite soldiers but he lost anyway.

Not because of a smarter enemy, but rather because he couldn't feed his army.

600,000 soldiers marched into Russia in 1812, but supply lines collapsed, food vanished, and horses died.

And by winter, the Grande Armée had been reduced to starving men chewing leather and stumbling barefoot through the snow, with only 100,000 making it back.

​There is a famous military maxim that Napoleon himself popularized: "An army marches on its stomach."

It doesn’t matter how brilliant your strategy is, how elite your soldiers are, or how perfectly choreographed your battle plans look on paper.

If you can't solve the logistics problem, the strategy doesn't matter, and in 2026, the AI industry just discovered its stomach.

For three years, every lab, every startup, every VC pitch deck has been obsessing over the same question: who has the smartest model?

OpenAI vs. Anthropic, GPT-5 vs. Claude: Benchmark wars across reasoning, math, and coding.

The logistics problem, the chips needed to run these agents at scale, became an afterthought.

But now it's one of the only things that matters.

The army is brilliant and enormous, but it’s starving and not for the chip everyone’s been talking about.

The chip everyone forgot

For three years, the AI infrastructure story was perfectly simple.

GPUs are king, Nvidia dominates, and the playbook became simple: buy, rent, and build everything around the latest chips like Blackwell.

The world made sense but there is a new villain in this story, and it is one of the most profoundly unsexy pieces of hardware ever invented.

The CPU.

Stick with me here, because I know CPU shortage sounds like something your IT guy mumbles about before asking you to restart your computer.

But this is actually the story.

In the early era of AI, ChatGPT answers questions, basic chatbots, and image generators, and the GPU does the heavy thinking.

It handles the matrix math, the model inference, all the exciting stuff, and the CPU sat in the corner like a middle manager at a startup.

Technically present but not really needed, handling maybe 5-10% of the total workload and then going back to its crossword puzzle.

Look at the left bar in this chart from TrendForce.

That was traditional AI, CPUs in the background, GPUs doing nearly everything, and an entire industry built around that assumption.

Now look at the right bar.

That's agentic AI, the one everyone is rushing to build right now.

The CPU just went from a footnote to half the entire operation.

Here is why that happens, and it's actually pretty intuitive once you see it.

A chatbot has one job: you ask, it thinks, it answers, then goes back to sleep, simple, clean, and GPU-driven.

An AI agent is a completely different animal, it doesn’t just answer questions, it does things: browsing the web, running code, checking your calendar, calling APIs, and remembering past context.

And it coordinates with other agents in parallel and runs in a continuous loop, often without you even looking.

Every single one of those steps, the tool calls, the memory lookups, the routing, the orchestration runs on the CPU.

The GPU is still doing the thinking part, but the CPU is now doing everything around the thinking, and it turns out everything around the thinking is roughly half the entire job.

Jensen Huang put a number on it at GTC 2026 that made everyone in the room do a double-take: agentic AI can consume one million times more tokens than a standard chatbot prompt.

Which means you don't just need a few more CPUs than you planned for.

You need a fundamentally different ratio of CPUs to GPUs than the infrastructure the entire industry just spent three years building.

Global AI token usage doubled from 6.4T to 13T tokens in just six weeks. 

IDC projects enterprise AI agent usage increases tenfold by 2027 and Agent-related API call loads are rising a thousandfold.

The army kept ordering more cannons and it turns out that what it actually needed was food.

Napoleon marched into Russia without enough food because he assumed the campaign would be over before winter hit.

The AI industry built its entire infrastructure without enough CPUs because it underestimated how many agents would actually need them.

The numbers that should terrify Intel

Here is where the story gets genuinely interesting.

Intel's own CFO David Zinsner stepped onto the Q4 earnings call and admitted publicly, on the record, that server CPU demand had caught the chipmaker off guard.

The company that has manufactured CPUs for 50 years.

The company built around server processors was caught off guard by demand for its own core product.

Revenue from Intel's Data Center and AI segment surged 9% year-over-year to $4.7B.

Sounds great, except Zinsner followed it up with: "Revenue would have been meaningfully higher if we had more supply."

They left money on the table because Intel couldn't make the chips fast enough.

So Intel did the only thing it could do in the short term: it started cannibalizing its own consumer business.

They diverted wafer capacity away from laptop and desktop chips to make more server CPUs, a decision that caused a 7% decline in their Client Computing Group revenue.

They robbed Peter to pay Paul, and Paul is still hungry.

In China, wait times for certain server CPUs have stretched to six months.

Some vendors are seeing 8-10 week delivery timelines just for standard enterprise chips and server CPU costs in China are up over 10% and climbing.

Even Amazon’s AWS can’t keep up, customers are already trying to lock up all of its Graviton CPU capacity for 2026.

When buyers are asking to reserve an entire year of supply from the largest cloud provider in the world, you’re looking at real scarcity.

THE GREAT CPU FAMINE (P2)

Now, here is where it gets fun because when Intel stumbles, somebody else picks up the fumble.

And that somebody is AMD.

While Intel has been scrambling to reallocate wafer capacity, apologizing on earnings calls, and watching its consumer business bleed, AMD has been quietly eating its lunch in the server room.

AMD's server CPU market share hit 28.8% in Q4 2025, up from 25.2% just a year earlier.

That might not sound dramatic until you remember that AMD was at roughly 3% server CPU market share in 2017.

They have gone from a rounding error to nearly one-third of the entire market in eight years, and the trajectory is still pointing up.

The specific product driving this is AMD's 5th Generation EPYC chips, code name Turin, which accounted for more than 50% of AMD's server revenues for the first time ever in Q4 2025.

AMD grew server CPU shipments at more than triple the seasonal average that quarter.

For context on how meaningful this is: Intel still holds 71% server CPU market share today.

AMD is at 28.8%, but the direction of travel is unmistakable.

Every quarter that Intel struggles with supply, AMD gains ground that it historically does not give back, and AMD is not standing still.

The company is simultaneously positioned in GPU territory with its MI300X accelerators, giving it a rare double presence in both the CPU and GPU sides of the AI infrastructure stack.

One company covering both gaps in a shortage that spans both chips.

We saw this one early and took a position in AMD before most people started connecting the dots.

The signs were right there, share gains, product strength, and Intel clearly on the back foot, and we’re already up over +20% on the position.

The next dominance shift

But here is the plot twist that changes the decade-long picture.

Even as AMD gains ground on Intel, both of them are facing a third threat that makes their competition with each other look almost quaint.

X86, the architecture Intel created and AMD licensed, the one that has powered nearly every server, laptop, and desktop computer on earth for 40 years, is losing to a fundamentally different approach.

ARM.

By 2025, half of all new compute capacity shipped to the big hyperscalers was ARM-based because ARM chips deliver dramatically better performance per watt of electricity consumed.

And when data center electricity bills are the binding constraint on your entire business, when you are literally buying nuclear power plants just to keep your chips running, performance per watt stops being a spec sheet footnote and becomes the single most important number in your P&L.

AWS built its ARM chip, Graviton, in 2019, and everyone thought it was a curiosity.

Today, it runs inside 98% of their top 1,000 customers, and AWS added 50% of all new CPU capacity over the past two years on ARM, surpassing both Intel and AMD combined in new capacity additions.

Atlassian disclosed at AWS re: Invent 2025 that they went from 5% to over 30% Graviton adoption in a single year because the cost savings were too obvious to ignore.

And then, in March 2026, Arm Holdings, the British chip design company that licenses its architecture to virtually everyone, did something it had genuinely never done in its 35-year history.

It built a chip.

The Arm AGI CPU is their first-ever production silicon, designed from scratch for agentic AI data centers.

More than double the performance per rack versus x86 systems, custom chip development timelines compressed from 3-5 years down to 6-9 months.

Customers already committed: OpenAI, Meta, SAP, Cerebras, Cloudflare, SK Telecom.

CEO Rene Haas is projecting CPU demand to quadruple in relation to agentic AI growth.

He has set a $25B revenue target for 2031, with $15B of that coming from in-house chip sales.

The response from Intel and AMD was the most telling signal of all.

In late 2024, they formed the x86 Ecosystem Advisory Group, a formal defensive alliance. Microsoft, Alphabet, Meta, and Broadcom are on the founding board.

Two old rivals joining hands because a third party is winning.

If you needed one image to understand how completely the old order is fracturing, that is it.

NVIDIA’s pivot nobody noticed

Here is the other move that matters, and it is not about GPUs.

NVIDIA, the undisputed king of AI chips for the past three years, watched the agentic shift coming and made a quiet but significant bet.

At GTC 2026, Jensen Huang unveiled the Vera CPU, Nvidia's own processor specifically designed for agentic AI orchestration.

The company that built its entire modern empire on graphics processors just decided it needs to be in the CPU business too.

Because in an agentic world, the orchestration layer, the part that coordinates all the tool calls, manages memory, and routes work between different AI systems, runs on CPUs.

NVIDIA now sells one CPU for every two GPUs in its Blackwell NVL72 configurations.

Their infrastructure leader told CNBC directly: "CPUs are increasingly becoming the bottleneck in scaling AI and agentic workflows."

When the GPU company starts selling CPUs because CPUs are the new bottleneck, you know the shift is real.

The winners and the losers

BITE-SIZED COOKIES FOR THE ROAD 🍪

Anthropic is turning down funding offers valuing it at $800B+. It’s holding off despite strong investor demand and rising revenue.

Snap is laying off 1,000 employees, about 16% of its workforce. The cuts aim to reduce costs as AI helps automate more work.

Amazon unveiled a slimmer, faster Fire TV Stick HD with Alexa+. It also opened preorders for Ember Artline TVs focused on design and AI features.

MILKY MEMES 🤣

ROADIE REVIEW OF THE DAY 🥛