GM. This is Milk Road AI, where we separate the AI spectacle from the businesses actually making the money.
Here’s what we’ve got for you today:
- ✍️ The hidden $2T engine of AI.
- 🎙️ The Milk Road AI Show: Why NeoClouds Could Become The Biggest AI Winners.
- 🍪 Anthropic’s $45B SpaceX compute deal.
Kalshi is a regulated prediction market where you can trade on real-world events. Special offer for Milk Road Readers: Trade $10 on Kalshi → Claim $10 for free.
Prices as of 10:30 a.m. ET.
THE $2T AI INFRASTRUCTURE GIANT
The most powerful man at Versailles wasn't the king but the man who controlled the water.
Versailles ran 2,400 fountains, but keeping them all running required a network of pipes, pumps, reservoirs, and engineers so complex that Louis XIV never let a single visitor see it working.
Source: Monsieur Feuilly Writes
Tourists came from across Europe to watch the water dance, writing about it in letters home, painting it, and talking about the spectacle for the rest of their lives.
But nobody wrote about the pipes.
The spectacle got the attention, but the infrastructure made the money.
Right now, the AI industry has a lot of fountains: NVIDIA’s H100, ChatGPT, Gemini, Grok, and every week brings a new model, benchmark, or demo that makes your jaw drop.
And everybody is watching the water dance.
But underneath all of it, quietly running 24 hours a day, is a $2T company called Broadcom.
It makes the chips that route data between GPU servers, the custom accelerators powering AI workloads at Google, Meta, and OpenAI, and the software running roughly 70% of the world’s enterprise servers.
It is the pipes under Versailles.
And while everyone else is busy watching the fountains, Broadcom is compounding at a rate that should make you put down whatever you're doing and pay attention.
What Broadcom actually built
To understand why this company matters, you need to know what it does, and I'll warn you: none of it sounds exciting.
Broadcom doesn’t design the AI model, run the data center, or own the consumer app.
What it has is three businesses that sit directly beneath everything happening in AI right now, and it has near-monopoly positions in all three.
The first business is custom AI chips.
Every major hyperscaler, Google, Meta, Amazon, Microsoft, OpenAI, and ByteDance is spending billions trying to escape NVIDIA.
The reason is pure math: NVIDIA’s GPUs are brilliant, but they’re built for everyone, which means they’re optimized for no one specifically.
When you’re deploying a million chips, that inefficiency gets expensive fast, and a chip custom-built for your exact model architecture, workload, and power constraints runs dramatically better than a general-purpose one.
So the hyperscalers went looking for someone who could build them a custom chip, one built specifically for their AI systems and co-designed over 18 to 24 months of engineering collaboration.
That someone is Broadcom.
Broadcom provides the silicon IP, the advanced packaging, and the co-design expertise that turns a hyperscaler's chip concept into something TSMC can actually manufacture at scale.
The customer designs the logic.
Broadcom provides the interconnect architecture and packaging, TSMC builds it and the result is up to 50% lower power consumption and over $2B in CapEx savings per 100,000-chip cluster versus buying NVIDIA GPUs off the shelf.
Google has been doing this with Broadcom since the original TPU, Meta followed, OpenAI just committed to a multi-year deal for 10 gigawatts of custom accelerators starting in 2026, and ByteDance is now in the mix too.
Broadcom currently holds roughly 60% of the AI server custom chip design partnership market, and the next closest competitor is Marvell, at around 10%.
The second business is Ethernet networking.
This one is even simpler and arguably even more dominant.
Every AI training cluster is essentially thousands of GPU servers that need to talk to each other constantly, passing enormous amounts of data back and forth at speeds that would have seemed impossible five years ago.
The chip that sits at the center of that conversation, routing all that traffic between servers, is called an Ethernet switch.
Broadcom makes the Ethernet switch that runs in roughly 80% of data centers on Earth.
Their current flagship, the Tomahawk 6, delivers 102.4 terabits per second of switching capacity in a single chip.
That's double the bandwidth of its predecessor, released just a couple of years ago.
It also natively integrates co-packaged optics, which solves the power and bandwidth bottleneck that becomes the real constraint when you're running AI training clusters at serious scale.
The industry is also moving away from NVIDIA's proprietary InfiniBand networking toward open Ethernet.
Every time a hyperscaler makes that switch, NVIDIA loses networking revenue, and Broadcom gains it.
Ethernet is rapidly becoming the dominant networking layer for large-scale AI clusters, with the market projected to exceed $30B by 2028 as InfiniBand growth begins to flatten.
Arista Networks, which buys Tomahawk chips and packages them into full data center switches, recently guided to $3.5B in AI fabrics revenue.
That number is a direct leading indicator of Tomahawk demand, Arista's growth is Broadcom's growth.
The AI networking market is projected to exceed $30B by 2028, growing at 38% annually.
Broadcom starts that race with 80% market share and a chip that nobody else can match in performance per watt.
Now their third business is VMware.
This is the one that sounds the least like an AI story, but is becoming one of the most important.
VMware's hypervisor software runs on approximately 70% of the world's enterprise servers.
If you work at a Fortune 500 company and your laptop connects to a corporate server for anything, there is a very high probability that the server is running VMware underneath.
It is the operating system of corporate IT globally, and it has been for two decades.
Broadcom acquired it for $69B in 2023, which looked expensive at the time, but the integration has been aggressive.
Broadcom eliminated perpetual licensing, forced customers onto subscription bundles, and raised prices by 800 to 1,500% in some cases.
The backlash was loud, European regulators opened an investigation, and large enterprises publicly threatened to migrate.
Then, over 90% of VMware's top 10,000 customers signed the new contracts anyway.
Because replacing VMware means redesigning your entire virtualization layer, rebuilding your software stack, retraining your IT teams, and accepting downtime risk on infrastructure that runs your business.
The cost of leaving is worse than the cost of staying, even at the new prices.
VMware generated $7.6B in revenue before the acquisition, but Broadcom’s software revenue reached $27B in FY2025 after integrating the business.
But the longer-term play is bigger than the price hike story.
Broadcom is quietly positioning VMware Cloud Foundation as the private enterprise AI platform for Fortune 500 companies that can't put sensitive workloads into public cloud.
Combine Broadcom's AI networking hardware with VMware's software-defined infrastructure, and you have a complete on-premise AI stack that no other single vendor can offer.
Chips, switches, storage controllers, and the software layer, all from one company.
The numbers that explain everything
In FY2025, Broadcom generated $63.9B in revenue and $26.9B in free cash flow.
That free cash flow conversion is staggering.
They are turning 42 cents of every revenue dollar into actual cash.
For context, Apple, one of the most celebrated cash machines in business history, typically converts around 25 cents.
But the number that changes everything is the AI revenue line.
AI revenue was roughly $3-4B in FY2023 but then it hit $12.2B in FY2024.
It hit $20-21B in FY2025, and Q1 FY2026 alone came in at $8.4B, up 106% year over year.
JPMorgan estimates Broadcom will do $50B+ in AI revenue in FY2026.
Management has signaled a $100B target by FY2027, and that's not just optimistic forecasting from someone trying to justify a price target.
It's backed by a $73B disclosed backlog.
That's roughly 1.5 years of forward AI revenue already committed, contracted, and waiting to ship.
Most semiconductor companies would kill for one quarter of that kind of visibility.
CRYPTO’S THIRD PMF MOMENT IS HERE
PMF = Product Market Fit
Until now, crypto has really had two major PMF moments: 24/7 trading and stablecoins.
Now a third one may be emerging with Kalshi: Prediction Markets.
Kalshi is a regulated prediction market exchange where people trade on the outcomes of real-world events (elections, inflation, sports, crypto prices and more).
And their growth has been massive:
- Recently raised $1B at a $22B valuation
- Institutional trading volume is up 800% in six months.
- Annualized volume is somewhere around $178B.
Kalshi is giving Milk Road readers $10 free when you trade $10.
Trade $10 → Claim $10 on Kalshi.
THE $2T AI INFRASTRUCTURE GIANT (P2)
In October 2025, Sam Altman called Broadcom "a critical step in building the infrastructure needed to unlock AI's potential."
He was announcing a multi-year partnership to design and deploy 10 gigawatts of custom AI accelerators, with production starting in H2 2026 through 2029.
That's the equivalent electricity consumption of over 8M U.S. households, committed to running Broadcom-designed chips.
But here's the part that didn't make the headlines, the entire system runs on Broadcom's Ethernet networking instead of NVIDIA's InfiniBand.
NVIDIA doesn’t just lose the compute revenue from OpenAI’s next infrastructure buildout, it loses the networking revenue from it too.
Google, Meta, and now OpenAI, the three most important AI spenders on earth have all called the same company to build their custom chips.
That's not a coincidence but rather that’s the moat.
And the man sitting at the center of all of it is Hock Tan.
He is not a product visionary, a disruptor, or a guy giving TED Talks about the future of humanity.
What he has is a playbook, and it works every single time.
Find a company with a near-monopoly position in an essential market where customers cannot easily switch, then acquire it, strip out everything that doesn’t generate cash, and raise prices to what the market will bear.
Convert the revenue into recurring subscriptions, maximize free cash flow, and repeat the playbook again and again.
He ran this playbook on CA Technologies in 2018, Symantec Enterprise in 2019, and most aggressively on VMware.
And Hock Tan's FY2025 compensation was $205M, of which $202M is equity tied directly to Broadcom hitting $120B in AI revenue by 2030.
That alignment matters. When a CEO's personal financial outcome is directly linked to a specific number, you tend to believe he's going to try pretty hard to hit it.
Why can't the competition catch up
The easy question is: if Broadcom is this good, why doesn't someone just build a competitor?
The answer is that by the time you asked that question, you were already about 15 years too late.
Broadcom's Tomahawk Ethernet switching chip represents over 15 years of continuous iteration.
The engineering knowledge embedded in that chip cannot be acquired quickly.
It cannot be hired away in a year or reverse-engineered without rebuilding the entire institutional knowledge base that created it.
The custom chip business is even harder to replicate.
Every ASIC generation Broadcom co-designs with a hyperscaler adds proprietary knowledge about that customer's architecture and roadmap.
That knowledge is non-transferable, and the moat deepens with every chip generation.
Marvell is the closest competitor in custom silicon, with roughly 10% Ethernet switching share versus Broadcom’s 80%, but while they’re growing, they’re not catching up.
NVIDIA dominates merchant GPUs but is actually losing networking share as the industry shifts away from its proprietary InfiniBand toward open Ethernet, and that shift is a structural tailwind for Broadcom.
Intel, after years of trying, has essentially exited the competitive custom silicon race.
Gaudi, their AI accelerator, is gaining minimal traction, and their foundry ambitions are years behind TSMC.
The honest part
None of this comes free.
Broadcom is currently trading at roughly 78x trailing earnings and around $414 per share.
For the stock to generate meaningful upside from current levels, the $100B FY2027 AI revenue target needs to materialize, and the market has already priced in a lot of good news.
There are real risks worth taking seriously.
Each custom chip generation is essentially a binary outcome.
If a hyperscaler’s ASIC underperforms NVIDIA’s competing GPUs, they can pivot back to merchant silicon fast.
Revenue is not a smooth curve because every chip generation must succeed technically, and not every generation will.
There’s also a scenario, not the base case but not zero, where hyperscaler AI spending pauses because the return on investment disappoints at scale. If the AI buildout takes a breather, Broadcom will feel it harder than almost anything else in the sector.
So here’s the honest framing: Broadcom is not cheap, it is excellent, and those are two very different things when deciding whether to buy the stock.
The $73B backlog de-risks the near term in a way most semiconductor companies simply cannot offer.
The free cash flow engine means Broadcom can pay down roughly $62B in VMware acquisition debt in under 2.5 years at current rates while still raising the dividend.
And if you want to see whether I’m buying Broadcom or not, come join Milk Road PRO.
Our PRO members get alerted first and also get access to the rest of my positions, watchlists, highest conviction ideas, and deep dives before the rest of the market catches on.
And have I mentioned the stock I wrote about last week is already up roughly 40% since that report? That’s the kind of move you’re missing out on.
Alright, that’s it for this edition of Milk Road AI. We want to hear from you.
Is Broadcom the most underrated AI infrastructure play?
- Yes: $73B backlog, 80% Ethernet share, custom chips for Google, Meta, and OpenAI. It's obvious in hindsight.
- No: 78x earnings is priced for perfection. One bad quarter and it gets ugly fast.
- Never heard of them: And yet you're running on their infrastructure right now.
Real Finance Blockchain is an EVM-compatible L1 that is built specifically for RWA tokenization. Read more about Real Finance Blockchain here.
Arc is an open Layer-1 blockchain built for real-world financial applications. Public testnet is live, App Kits are here, and mainnet is getting closer. Start building on Arc.
Your carrier is selling your data. Use code MILKROAD for 33% off your first 6 months.
BITE-SIZED COOKIES FOR THE ROAD 🍪
Special offer for Milk Road readers. Kalshi is giving Milk Road readers $10 free when you trade $10 on their platform.*
Google's Gemini Omni Flash is now live, turn any text, image, or audio into video. Available today to all Gemini Plus, Pro, and Ultra subscribers globally.
Anthropic is paying SpaceX $1.25B every month, a $45B compute deal through 2029. The deal was buried in SpaceX's IPO S-1, one of the largest AI infrastructure contracts ever.
Anthropic projects its first profit ever: $10.9B in Q2 revenue and $559M operating income. Revenue more than doubled from Q1 and Claude Code alone hit $1B annualized in six months.
*this is sponsored content.
MILKY MEMES 🤣
ROADIE REVIEW OF THE DAY 🥛
