GM. This is Milk Road PRO, serving the AI investment cake one layer at a time so you know exactly which part tastes the best.
Big news: Bot and agent traffic just surpassed human traffic on the internet for the first time: 57.4% machine, 42.6% human.
It’s clear that AI adoption is accelerating, and the model companies behind it are scaling faster than any software category in history. Anthropic filed its S-1. SpaceX starts trading June 12. OpenAI targets Sept. This is the biggest IPO wave since the internet.
But… only 5% of the GPUs already shipped are actually running. The other 95% sit idle, waiting on power and data center capacity that doesn't exist yet.
That gap is my thesis, which means I’m skipping the IPOs.
Demand is real, but the physical layer underneath it: power, data centers, and everything inside the rack, can't scale fast enough to keep up. The physical bottleneck caps how fast every layer above it can grow, and that's where both the opportunity and the risk sit.
I think it's better to focus on players that squeeze more compute out of every watt, especially while everyone is distracted by the gravity of the moment.
And before we continue: if you want to see all my trades in real-time, and exactly how I’m planning to play this thesis as the market takes a step back, you can join me in Milk Road PRO.
It’s only a buck for a quick seven-day trial, which you can cancel anytime. And you’ll see the full portfolio, along with the moves from our other 4 analysts.
How to think about AI investing
Every issue runs through one framework: the 5-Layer AI Cake that NVDA's Jensen Huang uses to map how value flows through the AI economy (see below).
Bottom to top, energy powers the chips, chips sit inside the infrastructure, infrastructure runs the models, and models drive the applications where the economic value finally shows up.
Each layer pulls demand from the one above. When an application takes off, it needs more models, which need more infrastructure, more chips, and more power.
We use the Cake to stay disciplined. For every development, two questions: which layer does this hit, and what does it mean for the layers around it?
Every position in the portfolio maps to one or more layers, and each issue walks the Cake bottom-up: what moved, what it means, and how it connects to the portfolio.
Source: Milk Road PRO
Here's what we're covering today:
- 🤖 Agents take over enterprise.
- 🏦 Model layer: the IPO wave.
- 🏗️ Infrastructure falls behind.
- 🧠 Chips: compute per watt wins.
- ⚡ Energy: off-grid power goes structural.
And of course, an update on my PRO portfolio and what I’m watching next.
AGENTS TAKE OVER ENTERPRISE
AI adoption started with individuals: consumers on ChatGPT, developers on Claude, creators on image generators. That phase is maturing. The shift now is into enterprise, where companies are moving from testing AI tools to embedding agents into core workflows.
Microsoft held its annual developer conference, Build, in late May, and it confirmed what's been building for months: agents are becoming the operating system.
They are rebuilding Windows as an agent runtime. Its "Agent Computer" lets agents write code, run commands, operate cloud PCs, and work inside policy-defined boundaries, all at the OS level.
Project Solara adds agent-first hardware (desk devices, wearable badges) built with MediaTek and Qualcomm, and Aion 1.0, a 14B-parameter reasoning model, ships in-box with Windows for local agentic loops with no round trip to the cloud.
NVIDIA unveiled RTX Spark at the same event: a one-petaflop chip built for on-device agents, pulling compute out of data centers and onto desks. Demand is expanding in both directions at once, more in the cloud and more on the device.
AI used to be a cloud product you logged into. Built into the operating system, it becomes infrastructure: every enterprise device a potential compute consumer, every workflow a potential AI workflow.
A governance layer is forming alongside it. Coralogix raised $200M at a $1.6B valuation in early June to monitor AI agents (hallucination detection, token cost tracking, security breach flagging), and over half its 5,000+ enterprise customers already run AI-driven incident investigation. As agents scale, someone has to watch them.
What this means
Enterprise integration opens three new dynamics:
Firstly, more stable and recurring compute demand as entire workstreams shift from human to agentic employees.
That's the following equation: More agents = more inference = more model usage = more data center demand = more chips = more power.
Secondly, the growing demand for AI governance, observability, and orchestration, which creates new opportunities on the application layer.
Finally, as we are seeing the growing adoption of agentic AI in enterprises, we should see elevated return on investment in downstream industries, e.g., healthcare or finance.
MODEL LAYER: THE IPO WAVE
The model companies powering this wave are no longer startups. They're becoming the largest companies in tech.
Anthropic filed its confidential S-1 on June 1, valued at $965B after a $65B Series H, with a revenue run-rate around $47B, up from roughly $10B a year ago.
Roughly 80% comes from enterprise API contracts and developer tools, not consumer subscriptions. Eight of the Fortune 10 are Claude customers, and over 1,000 companies spend more than $1M a year.
Anthropic is ramping revenue faster than any company at scale ever has: the prior record holder, Google, took 2002 to 2005 to climb from $400M to $6B. CEO Dario Amodei said the trajectory has outpaced internal forecasts by 8x.
Across the whole model layer, the numbers get larger. OpenAI and Anthropic combined have already passed Workday, ServiceNow, and Adobe in revenue, and on a run-rate basis they're now bigger than Google Cloud and Azure, with some estimates putting them past AWS by year-end.
No category in tech history has scaled this fast.
Source: Milk Road PRO (adapted from Coatue)
The catch for investors: run-rate is not the same thing as durable, profitable revenue.
Anthropic committed up to $100B to AWS in exchange for Amazon's investment, with a parallel arrangement at Google. Part of every lab's revenue is the cloud provider paying the lab to consume compute, and vice versa, so both numbers land in both top lines.
A $40B run-rate with low retention and one with high retention look identical from the outside, but one is worth a fraction of the other. Until the S-1 breaks it out, nobody can tell which is which.
That's why the S-1 matters: the market needs to see real gross margins, customer concentration, compute commitments, CapEx, retention, pricing power, and cash burn.
Right now the model layer is priced on growth. Public markets will judge it on economics: either these companies sustain software-like margins at infrastructure-like scale, or the market wakes up to growth bought with enormous compute spend.
Either outcome would reshape the actual opportunity in the market.
Chinese open-source adds more pressure. While U.S. labs race to list at trillion-dollar valuations, Chinese models are quietly taking share.
Alibaba's Qwen passed Meta's Llama in cumulative Hugging Face downloads. Andreessen Horowitz estimates 80% of startups on open-source models now run Chinese base models like DeepSeek and Kimi, which match U.S. frontier results at a fraction of the cost.
This compresses model-layer margins. As the gap keeps closing, labs either absorb the hit or push hard on efficiency: smarter architectures, smaller models that perform comparably, and relief on their highest-cost items (e.g., memory).
At the same time, they lean harder on the components that buy efficiency: advanced materials, power semiconductors, networking.
Either way, the money flows down into the physical supply, just maybe not as much into the places (e.g., memory) the market is focused on right now.
SpaceX adds another layer.
It filed its updated prospectus on June 3: 555.6M shares at $135 each, a $75B target raise, a $1.77T implied market cap. The previous record IPO was Saudi Aramco at $29.4B. SpaceX is aiming for more than double that.
What makes it an AI story is xAI (Grok) and compute infrastructure.
On June 5, Google signed a compute deal worth $920M a month through June 2029 for roughly 110,000 GPUs at Colossus. Combined with Anthropic's deal ($1.25B a month), SpaceX is pulling in over $2B a month in compute revenue, and xAI has moved its own training to Colossus 2 while monetizing the older cluster.
The build-out speed keeps compressing too: the first Colossus build took 150 days, the second 90, the third 60. They are bucking the bottleneck trend.
The reality check for investors: the IPO sells only new shares, so no insider is selling on June 12. By the time retail gets in, the early risk has already paid off.
Founders Fund's $20M from 2008 is worth roughly $182B. Google's $900M from 2015 is north of $120B. The public is paying for the bet after it has already won.
And no insider can sell until the 180-day lock-up expires in Dec, when Founders Fund, Fidelity, Google, and Sequoia can finally offload positions worth tens or hundreds of billions.
Even partial sales would put billions of new supply on the market. And it will be the same scenario for other big ones coming.
Source: Milk Road PRO (adapted from Coatue)
INFRASTRUCTURE FALLS BEHIND
As we all know, the demand for AI infrastructure is outpacing what can physically be built. And it's worth actually inspecting some of the numbers behind this narrative. For example:
- NVIDIA partnered with IREN to deploy up to 5 GW of AI infrastructure globally, with Texas as the flagship site for its DSX AI factory architecture.
- TotalEnergies and Dell are deploying Pangea 5, a supercomputer that multiplies computing power by 6x ($100M investment).
- In Europe, Ardian and Verne announced a €5B next-gen digital infrastructure campus in Île-de-France with 500 MW of capacity.
Jensen Huang said at GTC Taipei that a 1 GW AI factory now costs $50-60B, heading toward $80-100B. Goldman Sachs projects $8T in total AI CapEx over the next six years.
The cost is going up because these facilities are fundamentally different from what came before. Higher-voltage distribution (800V DC), liquid cooling, and on-site power generation are becoming baseline requirements.
Operators are designing integrated systems, not installing racks. The complexity and cost per megawatt are increasing with every generation of GPU. That means fewer companies can afford to build at scale, and the ones that can take longer to do it.
JPMorgan's analysis using satellite imagery shows how the buildout is falling behind. Over 60% of data center capacity planned for completion in 2027 has not yet begun construction.
An additional 7% has been delayed. Among projects slated for 2027, only 6.3 GW are actually under construction compared to 21.5 GW announced.
Next to rising costs, the physical friction points are stacking up. Grid interconnection queues in the U.S. now average three to five years. Large power transformers have 18-24 month lead times and limited global manufacturing capacity.
Communities near data-center-dense regions are pushing back on noise, water usage, and energy costs. Virginia, the largest data center market in the world, is actively revising permitting rules. Multiple projects have been delayed or downsized due to local opposition.
This is the fundamental asymmetry of the AI buildout.
Software scales in weeks. A new model can go from training to deployment in months. But the physical infrastructure underneath it (the land, the concrete, the materials, the grid connection, the cooling system) moves on construction timelines.
So what?
For investors, this asymmetry is the thesis.
The gap between digital demand and physical supply is widening. The physical layer caps how fast the layers above it can grow.
Model companies can sign all the enterprise deals they want. If there's no power to run the GPUs, that revenue hits a ceiling. The physical bottleneck doesn't just create opportunity in infrastructure, it rather creates risk in everything above it.
So the investment implication is straightforward.
The outside world sets a hard cap on how much energy and physical capacity is available and that cap isn't changing anytime soon. So the only lever left is efficiency inside the data center.
If you can't get more power, you have to extract more compute from every watt you already have. These are the efficiency trades inside the rack, and they're becoming the central investment thesis.
CHIPS: COMPUTE PER WATT IS THE METRIC THAT MATTERS
If you can't build fast enough and you can't get enough power, the only way through is efficiency.
The traditional power distribution chain loses roughly 17% of energy before it reaches the GPU. Each conversion step wastes power as heat, which requires more cooling, which uses more power. It compounds.
800V DC architecture eliminates multiple conversion stages, improving end-to-end efficiency from roughly 83% to over 92% and reducing copper usage by 45%.
This changes what matters inside the rack.
Vera Rubin and why it matters
NVIDIA's next-generation AI platform entered full production. First shipments go to AWS, Google Cloud, Microsoft, and Oracle this summer. It delivers 2.5-5x the inference performance of the current Blackwell generation. Vera Rubin Ultra is expected in late 2027.
Vera Rubin runs on NVIDIA's Kyber rack architecture, which uses 800V DC power distribution. In the Hopper era, a single rack consumed roughly 40kW. Blackwell pushed that to 120kW. Vera Rubin racks will reach 600kW to 1MW. That's a 25x increase in rack power density within three years.
With physical constraints tightening from the outside, the differentiator becomes how efficiently each watt gets converted into useful compute.
Power semiconductors manage the voltage conversion.
Optical fiber handles data movement at a higher density.
Advanced materials provide insulation against the heat and electrical stress that 800V DC and 3D chip packaging create.
These components are becoming as important as the GPU itself.
Jensen Huang has taken strategic positions in each of these bottleneck companies:
- Corning & Lumentum: Confirmed share purchases for fiber and optical transceivers.
- Infineon: A strategic collaboration for power semiconductors for 800V DC.
- Marvell: A public endorsement that moved the stock 34% in a day (optical).
- Qnity: An R&D collaboration on next-gen materials for 800V DC.
His investment activity tracks the bottleneck map in real time.
Vera Rubin entering production means the 800V DC thesis is no longer theoretical. Every hyperscaler building these racks needs 800V DC power delivery, power semiconductors, optical fiber at higher density, and advanced materials.
ENERGY: POWER IS THE BINDING CONSTRAINT
At the bottom of the stack sits the most fundamental limit: Energy.
The DOE launched Agora in late May. It's a test platform that simulates what happens when hyperscale AI campuses connect to the grid.
The bottleneck is no longer just whether enough power exists, but whether the grid can absorb the volatility of AI demand. AI workloads spike and crash in ways the legacy grid wasn’t designed to manage.
By creating Agora, the U.S. government is treating AI power demand not as a data-center problem, but as a national infrastructure challenge.
Battery storage is becoming part of the solution to it. Large-scale battery systems can buffer the spikes created by AI workloads, smooth the load, and buy time for baseload power to catch up.
Meanwhile, Denmark's Energinet paused new grid connections for large consumers. Virginia is revising data center permitting rules.
These are early signals of what's coming everywhere. The physical world is pushing back at the regulatory level.
Only 5% (!) of deployed GPUs are currently online. The other 95% sit idle because the energy and data center infrastructure to power them doesn't exist yet.
The industry is 15-20% through an $8T buildout. The grid is already straining at 15%. As the application layer accelerates, that strain gets worse before it gets better.
On-grid solutions to overcome this are slow. Multi-year permitting, grid queue backlogs, community opposition. Off-grid solutions like solid oxide fuel cells can deploy in 55 to 90 days and bypass the grid entirely.
And because Vera Rubin racks run on 800V DC, the fit is direct: Bloom Energy fuel cells produce 800V DC natively, with no conversion losses.
A fuel cell behind the meter feeding directly into a Vera Rubin rack skips the entire traditional power chain presenting a much-needed fix for the energy gap.
Also, the cost dynamics are starting to shift. Utilities and regulators are increasingly asking large loads to shoulder a greater share of the infrastructure they require.
Alberta enacted a 2% levy on computer hardware for any grid-connected data center drawing 75+ MW. New Jersey now requires loads over 25 MW to pay for new generation capacity.
At least 18 U.S. states introduced bills creating special rate classes for large energy users. Across major markets, data centers are facing longer interconnection queues, rising transmission costs, and growing scrutiny over who pays for grid upgrades.
That's why behind-the-meter power is moving from backup plan to strategic priority.
Senator Cotton's DATA Act proposes exempting off-grid data centers from FERC oversight entirely. New Hampshire already signed similar legislation. Behind-the-meter doesn't eliminate every infrastructure dependency, but it gives operators more control over deployment timelines and reduces exposure to grid bottlenecks.
The policy direction is becoming increasingly clear: regulators want large new loads to bear more of the costs they impose on the system. Whether through special rate classes, transmission cost allocation, generation requirements, or new permitting frameworks, the era of abundant grid capacity is ending.
Off-grid power isn't just a niche bet. It could become a structural requirement.
HOW I’M BUILDING MY PORTFOLIO
So with all that in mind…here’s how I’m positioning.
My portfolio has 11 assets, including 3 crypto tokens. I also recently trimmed a huge winner of mine. If you want to see what they are, join Milk Road PRO for just a buck.
Alright, that's all I've got for you today. I hope you found it valuable!
PS: One change to flag
These reports will now be monthly instead of biweekly. You'll still see me every week on the Roll-Up, plus I'll be hosting a weekly deep dive focused on a specific asset, sector, or theme.
If you have something you really want to see covered, hit me up at @WhiteCollarExit on X.
Catch you in the next one. 🫡
