I’m a weird one. But if I had to describe myself I would posture that I’m someone who deeply enjoys understanding systems. I’m certainly not an AI expert. I’m an extreme hobbyist at best. But I do think my unique path has a perspective worth sharing.
It feels a little strange to say this out loud because AI is obviously real. I use it. Most people I know use it. I think AI will matter a great deal:
I do not think that the market understands what kind of story it is pricing as it relates to AI.
When most people think AI, they see software. They picture a chatbot. That is understandable. Most people encounter AI through a screen. They experience it as software, so they naturally think about it in software terms. That is not where my mind goes first.
I have spent a lot of time around Bitcoin mining at scale. I spent more hours than I can count talking with utilities, grid operators, and people in and around the TVA region about power, interconnection, curtailment, substations, and what large electrical demand actually looks like once it has to leave the spreadsheet and become real. I have also mined Bitcoin on CPUs, GPUs, FPGAs, and ASICs. I have seen what happens when a computational task moves from broad and flexible hardware toward narrower and more efficient hardware.
AI will not trace Bitcoin mining in a one to one way. The workloads are broader. The end markets are broader. But time in mining does leave you with particular instincts. It teaches you to take load seriously. It teaches you that efficiency matters more than the market usually assumes. It teaches you that broad compute often dominates first and specialized compute often dominates later. Most of all, it teaches you that physical infrastructure stories and software stories are very different things even when they get wrapped in the same narrative.
So when I hear AI, I see the system as a whole.
I see load. I see transformers, cooling, networking, land, financing, and years of capital commitments made against assumptions about the future that may or may not be true. I see the complex underneath the story.
Because of my perspective, I think the market is making several mistakes at once.
The first mistake is about the machine itself. Investors may be extrapolating the wrong compute curve. They may be assuming that the broad and power hungry layer of compute will remain the dominant path longer than it should.
The second mistake is about valuation. A large part of this buildout is still being discussed like software and valued like tech, even though it increasingly behaves like financed real estate and infrastructure.
The third mistake is about economic moats. AI may be strategically important and still trend toward tighter pricing margins, wider distribution, and thinner capture than the market currently expects.
If those misunderstandings begin to break at the same time, I think the unwind may be painful. I think it becomes a bailout and capital rotation story. I think a lot of money will eventually go looking again for something simpler, scarcer, and less dependent on someone else’s projections coming true. All roads come back to Bitcoin.
What is an LLM?
If you have never spent time with this topic, it is worth taking a little time to understand Markov chain math and why prediction systems work the way they do. Veritasium has a great explainer on this and the history lesson is very fun. Don’t get lost in the math. Understand that large language models are prediction machines.
Huge Buildouts
Being a consumer today means complexity is abstracted away as a service. That has a tendency to separate us from the process.
Compute is not abstract. Power is not abstract. Cooling is not abstract. Interconnection is not abstract. A transformer is not abstract. A substation is not abstract. These are all very real parts of the puzzle.
Every time an AI CEO casually chuffs the notion of building out several hundred megawatts of hyperscalers an unappreciated army of utility engineers quietly cries themself to sleep at night.
If you have spent time around industrial load, the word megawatt starts to mean something different. You stop hearing it the way an analyst hears it. You start to see feeder constraints, switchgear lead times, land procurement, cooling density, queue risk, and capital expenditures that have to become concrete, copper, and steel before they can become revenue.
McKinsey estimates that by 2030, data centers may require about $6.7 trillion globally to keep pace with compute demand, with roughly $5.2 trillion tied specifically to AI processing loads. Morgan Stanley argues that credit markets will be critical to meeting more than $1 trillion of data center financing needs through 2028.
The story here is very physical, more capital intensive financing dependent than people presume when they imagine AI as software disruption.
The Compute Curve May Not Be What Investors Think It Is
One of the important lessons Bitcoin mining taught me is that broad hardware tends to win early and narrower hardware tends to win later.
At first, flexibility matters most. Then efficiency starts to matter more. In Bitcoin, that path ran from CPU to GPU to FPGA to ASIC. Each step narrowed the field and improved the economics. When a workload matures, or when some layer inside the workload matures, economic pressure usually builds toward specialization.
I think the market may be drawing too straight a line from today’s GPU heavy AI world into the future. It is looking at current demand and assuming that the same broad compute layer will remain the economically dominant one for much longer than it probably will.
AI needs huge compute today. The important question is whether future inference will continue to require the same degree of broad, generalized, and power hungry compute that the market is currently capitalizing.
The research base here is more supportive of specialization than many people realize. A recent survey of LLM inference hardware explicitly compares CPUs, GPUs, FPGAs, ASICs, and other hardware paths. That by itself is revealing. The field is not behaving as though inference has permanently selected its final home.
There is also specific evidence that narrower paths can materially change the economics. The paper LUT-LLM designed an FPGA based language model inference system around memory based computation and reported 3.05 times to 6.60 times higher energy efficiency than selected GPUs in its tested setup, along with 1.10 times to 3.29 times faster generation speed in certain conditions. And these are very early improvements. The serving layer is still being re-architected in ways the current capex story may not be fully pricing or planning. And that serving layer is being attacked from several directions at once.
Quantization lowers the amount of precision needed for useful output. Speculative decoding lowers wasteful serial decode steps. Custom silicon reorganizes hardware around specific bottlenecks. Model routing can push simpler tasks toward cheaper paths. Edge and local deployment can peel some demand away from the most centralized clusters.
This is not just academic theory. The hyperscalers themselves are already building toward a more specialized future.
Microsoft’s Maia 200 is explicitly described as an inference accelerator built to improve the economics of AI token generation inside what Microsoft calls a heterogeneous AI infrastructure. Google’s eighth generation TPU line splits into TPU 8t for frontier training and TPU 8i for inference and reinforcement learning, with Google saying directly that the infrastructure requirements for pre training, post training, and real time serving have diverged. AWS says speculative decoding on Trainium can accelerate token generation by up to 3 times for decode heavy workloads by reducing serial steps and improving utilization.
The energy story is also very nuanced. A 2026 paper in Joule argues that many public estimates of inference energy are overstated by 4 times to 20 times because they rely on non production assumptions, and that model, software, and hardware improvements together could reduce per query energy by another 8 times to 20 times.
So my question is this: If we are massively scaling GPU data centers based on assumptions that might not hold true regarding the hardware end state and actual load, what happens when the industry pivots to narrower compute for significant portions of the load? How much sunk CAPEX is now stuck? Were the models generated able to recoup the investment?
Cheaper inference can produce more inference. Better efficiency can widen adoption. Longer reasoning traces can consume far more energy even while baseline serving gets cheaper. The future demand curve is not linear. It is not mechanically tied to today’s GPU intensity. Investors may be capitalizing a world in which the current compute path economics will allow.
Even if inference does not migrate down the stack as quickly as I suspect, impairment risk can still emerge at the GPU level because the hardware is improving fast enough to turn long-duration infrastructure bets into a shelf-life equation. Right now we are still deep inside compute scarcity, and hardware prices reflect that. I have spent my whole life around hardware, and I have never seen price dislocations like these. But supply meets demand. Manufacturing scales. Innovation compounds. And when that curve flips, a lot of assets financed at scarcity-era assumptions may suddenly look far more burdensome than they did on paper. Every economic incentive is pushing engineering progress forward at a blistering pace. The real question is whether these builds were financed for that reality, and whether the market can absorb the disruption if they were not.
That is the first misunderstanding.
The Market Is Still Talking About Tech While Financing Infrastructure
This is the deeper mistake. A lot of investors still seem to be asking whether AI is another dot com bubble. It is the obvious historical comparison. There is excitement. There are huge valuations. There is a race to own the future.
I have only recently become a hedge fund manager professionally so I spend a lot of time looking up to managers with deep experience. The piece that sharpened the financial piece for me most was The Second Derivative: Why No One Understands the AI Boom. The force of that essay is not that it says “bubble.” Lots of people can say bubble. Its force is structural. It argues that markets focus too much on the level of growth and the first derivative of growth while missing the second derivative, meaning whether growth itself is still accelerating or beginning to decelerate.
What does that mean? A system designed to be refinanced does not only need growth. It needs growth to keep arriving at a pace that satisfies the assumptions embedded inside the structure. Once that acceleration starts to fade, the machine can be in trouble even while the headline numbers still look healthy.
This is what made 2008 so widely misunderstood in real time. The clean story is that people were unable to pay their mortgages at a large enough rate and the market started to collapse. The harder story is that the refinancing treadmill began breaking before the public story had fully caught up with the sequence.
I think that frame is useful for AI than the cleaner dot com comparison. Dot com, in broad terms, was an equity multiple event. Too much optimism got priced into too many companies. The damage was real. But much of that damage lived in equity repricing.
What I see now looks very physical and much more financing dependent. It looks like land, power, cooling, racks, memory, networking, chip procurement, and increasingly large capital outlays financed through a mix of operating cash flow, debt markets, private infrastructure capital, and securitized structures. The AI buildout increasingly behaves like an infrastructure project with a software multiple attached to it.
The question becomes who is lending against what assumptions, which balance sheets are quietly carrying which risks, and what happens if future demand or future pricing evolves differently than the buildout requires?
A large part of the buildout may be functioning as a disguised loan book. Capex gets celebrated as confidence. Backlog gets celebrated as demand. But if those long dated commitments are being financed against assumptions that later weaken, then some of what the market is pricing as growth will date poorly.
A financing structure tied to hard assets can impair and likely will. It can strand capital. It can reveal, very late in the game, that the thing being valued was actually just a bet on continued acceleration.
That is the second misunderstanding.
Strategic Importance Is Being Confused With Durable Pricing Power
AI is clearly important. Some of the systems already available today are extraordinary. They are useful today and they will become much more useful.
That does not guarantee that every layer around them will enjoy software style margins the market is used to.
The competitive field is widening and the economics are already shifting under the surface. One of the clearest signals of that comes from open and open weight models, especially out of China.
A Stanford HAI issue brief, Beyond DeepSeek, puts the point plainly: Chinese made open weight models are now unavoidable in the global competitive AI landscape. The brief goes beyond DeepSeek and shows a deeper ecosystem that is prioritizing computational efficiency, flexible downstream deployment, and broad adoption.
Software companies are used to the economics of control. If a capable model can be downloaded, fine tuned, adapted, and deployed into private environments, then the premium attached to closed access and central dependence gets harder to defend across the market. Priced too high, and you incentivize independence which carries even more advantages above just cost. Closed models do not disappear. But the range of viable alternatives expands, and that alone changes the pricing conversation.
I think the user base is already splitting into distinct groups. There are power users and sovereignty minded builders who want control, lower dependence, and ownership over their data environment. There are enterprises that care about compliance, reliability, integration, and cost discipline. There are light consumers who mostly want cheap convenience and output that is good enough. It’s a race to the bottom on all fronts. This weakens the fantasy that every token will continue to be delivered through a highly centralized and permanently premium model.
The light consumer is used to software being cheap or free. The power user increasingly wants sovereignty. The enterprise buyer wants a procurement decision that still makes sense after the next cost decline. None of those groups are naturally preserving software style monopoly margins for the whole stack.
The pricing data supports this. A 2026 paper, Tiered Super-Moore’s Law, documents an approximately 600 fold decline in token prices and argues that the market has shifted from technology driven to competition driven price acceleration. Even more striking, the paper argues that software and architectural innovation are doing more of the cost reduction work than hardware advances alone.
That should matter a great deal for investors. A lot of people still seem to think about AI the way markets think about early software platforms: winner take most, high margins, high upside to scale compared to low capex requirements. I think a meaningful portion of AI may end up looking more like a utility layer than investors currently want to admit. I think AI business models become a strategically vital service whose economics get pushed steadily toward lower prices, wider access, harsher competition, and thinner capture across large parts of the stack.
That is the third misunderstanding.
When the Air Comes Out
This is where Bitcoin comes back into the frame.
I do not think Bitcoin and AI are direct competitors. Bitcoin is the best money humanity has and likely will ever create. AI is a sprawling mix of intelligence, infrastructure, services, and increasingly commoditized capability. They are different things.
They do compete for narrative oxygen and capital.
AI has clearly absorbed an enormous amount of both this cycle. The market has been trying to digest what AI is, who actually gets paid, and where all the excess capital sloshing around the system should go. Bitcoin may not really start to hum until the AI market is digested more fully.
The AI mania may be bigger than dot com. Oversupply and misallocation are real possibilities, and that the people rotating out of Bitcoin into AI may eventually discover that a technology can succeed while the current investment vehicles built around it can still fail badly.
“The first mouse gets killed. The second mouse gets the cheese.” — Luke Gromen
I think a lot of first mice are about to get killed.
AI is a debt financed, high expectation sector that may be especially vulnerable if tighter conditions, weaker growth, or a policy mistake force the market to price risk back into a system that has been behaving as though little can go wrong.
If AI is real but mispriced, and if the infrastructure and financing wrapped around it prove more fragile than the market expects, then a correction will not simply destroy capital. It will also dislodge capital. Some of that money will go looking for a new home. If the macro response to the correction includes rate reductions, easing, or more direct liquidity support, the question becomes where that capital flows next.
Bitcoin stands apart from the part of the market likely to be repriced. Every risk above is defined by counterparty failure. I think an AI bust might likely teach the value of carrying no counterparty risk.
Closing Thought
AI will likely prove every bit as important as the bulls say. It may reshape research, medicine, design, logistics, coding, and countless other fields. I suspect it will. The underlying technology is real.
I think the market is treating AI like a broad software miracle with tech style economics. What I see instead is a messier machine. I see hardware specialization attacking the assumption that today’s compute path is tomorrow’s compute path. I see a physical buildout that increasingly behaves like financed infrastructure. I see open weight competition applying pressure to pricing and control. I see a technology that may be profound while still proving economically harsh.
Bitcoin has taught me that a real thing can be mispriced. The question is, what side of the fence are you on right now?
Nate Slemp · BTC Strategy