GM. This is Milk Road PRO, where we map the next AI buildout bottlenecks before the market reprices them.
For most of the past couple of months, the market has asked one core question:
Do we still chase HBM, advanced packaging, and CPUs, or has that trade already gone too far?
Fair question.
Agentic AI changed the demand picture fast. Once AI moved from a simple “chat” into reasoning, coding, tool use, and multi-step workflows, investors realized these systems need far more inference, memory bandwidth, packaging capacity, and system-level compute than expected.
That is why names like Micron, SK Hynix, and AMD moved so hard.
And yet I don’t hold any of those three assets in my Milk Road PRO portfolio (even though my colleagues do).
Because stopping there misses the bigger point.
The AI buildout is not just one bottleneck. It is a sequence of bottlenecks. And based on the development outlined in this article, it's likely that the next will be in the following buckets.
Why agents change the bottleneck
In the chatbot phase, the system mostly needed enough compute to answer a question.
In the agentic phase, the system needs enough infrastructure to keep millions (eventually billions) of AI workers running continuously.
That is a very different workload.
An AI agent does not stop after one prompt and one answer. It reasons, calls tools, writes code, checks outputs, retries steps, waits for new information, and runs in the background.
That means more tokens, more traffic, more memory pressure, more power draw, and more uptime pressure per user task.
So the bottleneck expands. It is no longer only about the chip itself. It is about the full stack required to keep that chip running at scale.
How the buildout moves from here
This is not a “more memory and packaging forever” story.
These are real bottlenecks. But like NVIDIA before them, related stocks will move eventually from explosive repricing back toward more normal return profiles once the market understands the trade.
That does not mean the agentic AI buildout trade is finished. It means the bottleneck is moving.
Every AI phase creates a new constraint. Once one layer gets funded, scaled, and priced by the market, pressure shifts to the next layer of the stack.
That is the pattern investors need to understand.
As agents move from demos into persistent workflows, the constraint does not stay inside the chip package. It moves outward into the infrastructure that lets those chips run at scale.
The power stack gets interesting
AI data centers are turning into industrial power loads.
The chatbot phase was already energy-hungry. But the agentic phase is different because the workload becomes more persistent, more volatile, and more dense.
That creates two problems at once:
- Time to power: Can you get enough electricity to the site, and get the campus energized fast enough?
- Power density: Once power reaches the campus, can you deliver it efficiently into the next generation of ultra-dense AI racks?
That distinction matters because these are not the same problem - and they create different investment opportunities.
Time to power
The first energy bottleneck is already here.
AI campuses need massive amounts of electricity, but the grid was not built for this speed of demand growth. Utilities, interconnection queues, transformers, switchgear, and local permitting all move slower than hyperscaler CapEx plans.
That is the time-to-power trade. It has three main segments:
Behind-the-meter power
Instead of waiting 3-5 years on the public grid, the data center brings power to the site itself through fuel cells, gas engines, turbines, solar, batteries, or microgrid setups.
In plain English: The data center stops waiting for the grid and brings power to itself.
Bloom Energy is the cleanest example. It provides on-site fuel-cell power for AI campuses within 90 days.
GE Vernova and Siemens Energy have also benefited from the broader power scarcity trade, especially as gas turbines, grid equipment, and electrification became more strategic.
But this is no longer hidden. Bloom has been re-rated and the broader diversified names have also moved.
What could accelerate the trade even further is policy.
As utility bills for society in data-center-dense regions keep climbing, the government may be forced to enter the discussion.
One option could be a data-center levy that rewards self-generation.
If grid-pulled power gets taxed more heavily, while self-generated power gets a lower rate, the incentive flips.
AI power stops being a public grid burden and becomes a hyperscaler-funded behind-the-meter buildout.
That would directly benefit Bloom Energy and others as a fast, on-site power provider.
So, from here, upside depends mainly on proof: hyperscaler deals, repeatable deployments, and manufacturing scale. A major wild-card could be policy.
Battery energy storage systems (BESS)
Battery storage used to be mostly about backup. Now, it is becoming an active power-management tool.
AI workloads are not only bigger. They are more volatile.
More inference, denser clusters, and dynamic workload scheduling create sudden demand spikes that stress local grids.
Batteries help smooth that profile.
That is why a company like Fluence has started to matter more in the AI infrastructure discussion. It is one of the cleaner public storage integrator names.
Tesla Energy is also relevant through their Megapack battery, although Tesla is still mostly valued through robotaxis and robotics.
There is still potential for major upside in my view, especially in context of the solar energy buildout and the accompanying demand for batteries. However, the trade is also not completely unknown to the market, especially when it comes to Fluence and broader storage names.
Speciality grid hardware
Speciality grid hardware is about transformers, switchgear, substations, breakers or transfer switches.
Before a data center can run GPUs, someone has to safely connect, transform, switch, protect, and distribute electricity across the site.
Eaton, Schneider Electric, ABB, and Siemens Energy have already worked well because they sit directly in the path of data-center growth, electrification, grid upgrades, and AI power demand.
With the grid remaining a core bottleneck over the next few years, especially in data center dense regions in the U.S., there is still room for speciality grid hardware to be reframed as core AI infrastructure.
Power density: 800V DC architecture
The second energy bottleneck is more technical, but it may be the most underappreciated one.
Power density is about delivering electricity inside the data center and into the rack. This becomes much harder as AI systems move from normal server racks to ultra-dense rack-scale computers.
NVIDIA’s rack roadmap is the signal here.
Rubin starts the next upgrade cycle in 2026. But the bigger power-density shift likely comes in 2027, when Rubin Ultra / Kyber rack-scale systems make 800V DC commercially relevant.
The reason is simple: data-center AI racks are rapidly outgrowing legacy 48-54V supply schemes.
As per-rack power rises from roughly 0.1 MW toward 0.6-1 MW, older power architectures become inefficient.
And thus, we turn to the 800V DC.
800V DC means electricity is distributed at 800 volts direct current closer to the rack. Higher voltage allows the system to move the same amount of power with lower current.
That means less heat, lower losses, thinner cables, less copper, and better efficiency at rack scale.
In simple terms: 800V DC makes ultra-dense AI racks easier to power.
Why this trade is still early
AI servers already need more power components today. Every GPU rack needs power semiconductors, controllers, and converters to manage higher electrical loads and deliver stable power to the chips.
That is the current AI server-power trade.
Names like Infineon, Texas Instruments, STMicroelectronics, onsemi, and Navitas are getting some attention because AI increases the amount of power that needs to be converted, controlled, and protected.
But 800V DC is different and has not arrived yet.
That is the asymmetry.
Investors are not buying these names specifically because of 800V AI data-center power. If Rubin Ultra / Kyber makes higher-voltage DC distribution relevant, the market could start treating power semis as part of the next AI bottleneck.
Power semis beyond the rack
The 800V DC thesis does not stop at the rack and the power-semiconductor opportunity may be broader.
As AI campuses become massive electrical loads, power delivery also matters outside the rack: substations, solid-state transformers, load regulation, AC-to-DC conversion, and grid-edge power electronics.
That expands the thesis from rack-level conversion to campus-level power delivery.
The timing is important.
Power semis are not the tightest bottleneck today, partly because weakness in EVs and industrial end markets created spare capacity.
But that can change quickly if AI demand keeps compounding.
The setup looks like this:
- 2026 = narrative starts.
- 2027 = architecture and design-win trade.
- 2028 = revenue trade.
So, the 800V DC is still early. But if AI power demand keeps scaling, power semis may become important not only inside the rack, but across the full AI campus power chain.
The 800V DC second-order materials trade
800V DC is not only a power semiconductor story.
If racks can take much more power, they can pack more compute into the same footprint.
But that creates a second bottleneck: heat, electrical stress, insulation, signal integrity, and long-term reliability.
That is where thermal and speciality materials become more important. These products essentially make the higher-density AI rack survivable.
The market has already discovered the obvious version of this trade: liquid cooling. The market has also discovered advanced packaging.
But the broader speciality-materials layer is still less obvious.
Most investors are not yet looking one step further into dielectrics, adhesives, underfills, encapsulants, shielding materials, gap fillers, pads, gels, or coatings as core enablers of the AI infrastructure buildout.
Speciality materials are still mostly hidden inside broader chemicals, electronics materials, and industrial-supply buckets.
That is why the setup is interesting.
The market has started to price “AI needs cooling”, it has not fully priced “AI needs a broader reliability materials stack”.
How am I playing this?
Bloom Energy
I originally bought Bloom for the time-to-power trade. The market understands that part now - the stock is up over 100% since I bought it in February, and has been my top-performing holding.
The second leg of the thesis is what keeps me in the position.
Bloom's fuel-cell platform delivers onsite power within 90 days, and the architecture is 800V DC-native.
That aligns it with where NVIDIA's AI power roadmap is headed.
With that, Bloom could sit in a more strategic position than the market currently gives it credit for, and that is the piece I do not think is fully priced in.
That is where the NVIDIA angle gets even more interesting. Jensen Huang has already shown his willingness to invest across the AI infrastructure supply chain when a bottleneck becomes strategic.
If power becomes one of the key constraints for Rubin Ultra / Kyber-scale systems, a DC-native onsite-power platform like Bloom could become strategically relevant to NVIDIA’s broader ecosystem.
To be clear: this is not my base case. But it is a credible upside option that the market may not be pricing in.
IFNNY
I wrote about Infineon already in the last report, but let me give you a more nuanced view here.
As racks get denser, the bottleneck shifts toward power conversion, voltage regulation, protection, and efficiency.
Infineon sits directly in that layer through power semiconductors, controllers, and conversion technologies.
The numbers are already large enough to matter: Infineon reported AI data-center power revenue was more than $800M in FY2025, should reach around $1.75B in FY2026, and about $2.9B in FY2027.
Also important is the recent restructuring where Infineon is simplifying from four to three divisions to put clearer ownership behind its biggest system-level growth arenas.
Strategically, it separates Power Systems for AI data-center infrastructure and Edge Systems for edge AI, which aligns the org directly with the shift to AI power and physical AI.
This restructuring means that basically 50% of the company is focused exclusively on the AI buildout. The other 50% is focused on (legacy) automotive, which may give them a USP (Unique Selling Position) in the emerging robotaxis market.
The market has started to notice that, so Infineon is not a completely hidden AI name anymore. But I still think the broader thesis is underpriced:
- AI data-center power first.
- Physical AI (power) later.
That second layer matters because AI moving into the physical world means more humanoids, robotaxis, industrial robots, etc.
All of them need power control, motor control, sensing, and energy management.
That is Infineon’s long-term lane.
Their recent M&A activities also point toward that end: Infineon is buying ams OSRAM’s non-optical analog/mixed-signal sensor portfolio for $660M, adding ~$265M CY26e revenue and deeper sensor exposure in automotive, industrial, and medical markets.
Essentially, they are buying more sensing content, strengthening Infineon’s ability to sell system-level power, control, and sensing solutions into cars, robots, factories, and edge devices.
So, Infineon is a bet on a structural shift. The AI economy needs far more intelligent power electronics across data centers, robots, and edge devices.
Qnity
Qnity is a second-order AI infrastructure play on rising materials intensity (on my watchlist).
As AI hardware gets power denser (800V DC architecture ultimately) and harder to manufacture, more value shifts into the inputs that make chips, packages, boards, and systems work reliably.
That is where Qnity fits.
The company supplies semiconductor and electronics materials across two main buckets: Semiconductor Technologies and Interconnect Solutions.
That includes CMP pads and slurries, photoresists, plating chemistries, advanced packaging materials, films, laminates, thermal interface materials, gap fillers, coatings, EMI shielding, and reliability materials.
So the exposure is basically across the entire AI compute supply chain.
The thesis has three useful validation anchors.
First, Qnity is collaborating with NVIDIA on advanced semiconductor and electronics materials for AI and high-performance computing. This is a collaboration framed around materials innovation for packaging, interconnect, and thermal-management challenges.
Second, Qnity was added to Apple’s American Manufacturing Program. That validates its role as a U.S.-based materials partner for semiconductor production and AI-related technologies.
Third, Qnity is a focused DuPont spin-off. The business was previously buried inside a broader industrial portfolio. Now it has its own strategy, management incentives, disclosure, and investor base around semiconductor and interconnect materials.
That matters because we have already seen focused AI infrastructure spin-offs work well when the market realizes they sit on a specific bottleneck.
GE Vernova became the power/grid vehicle and SanDisk became a memory/storage vehicle.
So, as AI increases material content demand, Qnity may become the materials layer behind the AI hardware stack, helping chips get made, stacked, connected, cooled, shielded, and kept reliable over time.
Tactical setup: don’t chase the parabola
Now you are all bullish on those 3 names. To finish this article off, I want to outline why I'm not buying fully into those names (yet).
The thesis is intact, but the trade has moved fast.
Agentic AI pulled demand forward, hyperscalers ordered aggressively, and the market started pricing anything connected to AI scarcity (some things more than others).
That creates short-term tension.
Chips, memory, and packaging were ordered faster than data-center sites can be built, energized, cooled, and approved.
At the same time, many stocks tied to AI scarcity have already gone vertical, including some of the second-layer infrastructure names.
So this is not an “add everything now” setup.
If the AI trade digests, even the names I like in power, grid, and materials can get hit with the basket.
Therefore, I'm tactically separating thesis conviction from entry discipline.
For Bloom Energy, the time-to-power thesis remains strong. But after a major re-rating, I'm holding the core position and may trim into further strength rather than add aggressively.
For Infineon, I have started with a small position. It is my cleaner play on 800V DC, power conversion, and AI electrification, but I want to build it gradually, ideally on weakness.
For Qnity, I like the long-term materials-intensity setup, but it is more exposed to semi-cycle digestion.
So, I'm patient and would rather wait for a better entry.
Catch you in the next one. 🫡
- Vincent (@WhiteCollarExit)
AI-GENERATED PODCAST 🤖
We’ve turned this PRO report into an AI-generated podcast to make it even easier to digest. You'll find the audio player below. 👇️🎧️
The market is chasing the wrong layer
Disclaimer: This podcast was created using AI and is based on the research report above. While we've done our best to ensure accuracy, the audio may contain minor errors, technical glitches, or mispronunciations. Please note that this podcast provides an overview of the report and is not a comprehensive or definitive take on the topic.
