Oisin Moriarty

There's a strange asymmetry in how we think about the current AI boom. Everyone focuses on whether ChatGPT will keep getting smarter, whether OpenAI's valuation makes sense, whether enterprise adoption will accelerate. Almost nobody is asking: what happens to all the hardware if it doesn't?

Between 2023 and 2027, tech companies will deploy over $1 trillion into AI infrastructure. Data centers packed with Nvidia H100s and B100s. Purpose built cooling systems. Upgraded power grids. Networking equipment designed for moving massive training datasets. This is the largest infrastructure buildout in modern history, executed in under five years.

The bet is that large language models will continue scaling, that bigger models trained on more data with more compute will keep getting better, and that this improvement will generate enough economic value to justify the spending. But what if that bet fails? What if scaling laws break down, or LLM capabilities plateau, or enterprise customers decide the ROI isn't there?

I want to explore a different scenario. Not whether the infrastructure buildout is justified by LLMs, but whether it might accidentally enable something else entirely: a genuine revolution in materials discovery. The irony would be profound. We'd have built a trillion dollars of compute infrastructure chasing better chatbots, only to discover its real value was finding better magnets.

The Optimistic Case: How Compute Could Actually Transform Materials

Let's start with why materials discovery could genuinely benefit from scaled compute infrastructure in ways that go beyond current DFT limitations.

The core problem in materials science isn't just that Density Functional Theory is slow. It's that the search space is impossibly large and materials with commercially valuable properties are vanishingly rare. The periodic table has roughly 60 usable elements. Combine five of them in different ratios and you get over 5 million possible compositions. Add crystal structure variations, synthesis conditions, and operating environments, and you're talking billions of potential materials.

Current approaches sample this space inefficiently. Academic labs test maybe hundreds of compounds per year. Even high throughput experimental methods struggle to exceed thousands. You're searching for needles in a haystack the size of a galaxy.

Scaled compute infrastructure could change this in several ways. First, massive parallelization of quantum chemistry calculations. Instead of running one DFT calculation at a time on a university cluster, imagine running ten thousand simultaneously on a hyperscaler data center. You could systematically map regions of materials space that have never been explored, building comprehensive databases of calculated properties.

Second, ensemble methods. One DFT calculation might be unreliable, but what if you ran the same material through multiple different functionals, multiple basis sets, multiple simulation parameters, and used statistical methods to quantify uncertainty? This is computationally expensive, requiring perhaps 100x more compute per material, but it gives you confidence intervals on predictions. You'd know which materials are reliably predicted versus which ones need experimental validation.

Third, multi scale simulation. Real materials aren't perfect crystals. They have defects, grain boundaries, surface effects, all of which dramatically affect properties. Simulating these requires linking quantum calculations at the atomic scale with classical molecular dynamics at larger scales, with continuum mechanics at even larger scales. This is staggeringly compute intensive. A single multi scale simulation might require weeks on current hardware. With hyperscaler infrastructure, you could run hundreds in parallel, actually predicting how materials behave in realistic conditions rather than idealized environments.

Fourth, inverse design through massive search. Instead of forward prediction (here's a material, what are its properties), you do inverse design (here are the properties I want, find me materials). This requires exploring vast combinations and rapidly evaluating each one. Graph neural networks can accelerate this, but they still need substantial compute for both training and inference at scale. With access to tens of thousands of GPUs, you could search materials space far more thoroughly than currently possible.

Fifth, experimental data integration. The real breakthrough would be coupling scaled compute with high throughput experimental labs, creating a closed loop. Compute suggests candidates, robots synthesize them in parallel, automated characterization measures properties, results feed back into the models. This loop needs to run fast, with compute always ahead of the experimental pipeline, suggesting the next batch of materials to test. That requires substantial always on compute capacity.

Here's the critical insight: these approaches don't solve the fundamental limitations of DFT I discussed in my previous post. DFT is still inaccurate for complex systems, high temperatures, and magnetic properties. But with enough compute, you can work around these limitations through sheer volume. Run enough calculations with enough different methods, integrate enough experimental validation, search enough of the space, and eventually you find materials that work even if your models aren't perfect.

It's brute force, but brute force at sufficient scale becomes its own kind of intelligence.

The Infrastructure Reality

Now let's talk about what's actually being built and what it costs.

Microsoft is spending roughly $50 billion annually on capital expenditures, with the majority going to AI infrastructure. Google is at similar levels. Meta announced $37 to 40 billion in capex for 2024. Amazon's AWS infrastructure spending is comparable. Combined, we're looking at over $200 billion per year from just these four companies.

What does that buy? Data centers consume about 40 to 50 percent of capex, with the remainder split between servers, networking, and other equipment. A single large scale AI data center might cost $1 to 2 billion to build and equip. At current spending levels, the hyperscalers are bringing online perhaps 100 to 150 major facilities between now and 2027.

The chips are the most expensive component. Nvidia's H100 sells for around $30,000 to $40,000 per unit. A single training cluster for frontier models might contain 10,000 to 50,000 GPUs, representing $300 million to $2 billion just in chips. Microsoft reportedly has ordered over 100,000 H100s. The upcoming B100 will be more expensive, potentially $50,000 to $70,000 per chip.

These aren't generic compute resources. AI training infrastructure is highly specialized. High bandwidth interconnects between GPUs, specific cooling requirements, power delivery systems capable of handling megawatts per rack. You can't easily repurpose an AI training cluster to run traditional cloud workloads. The architecture is fundamentally different.

Power is becoming the binding constraint. A large AI data center can consume 100 to 300 megawatts. That's equivalent to a small city. Utilities are struggling to keep up. In some regions, data center construction is limited not by capital or space but by available electrical capacity. Companies are signing deals directly with power plants, even considering building their own nuclear reactors.

This infrastructure has a limited useful life. GPU performance improves rapidly, roughly doubling every 18 to 24 months. A chip that's state of the art today is obsolete in three to four years. Data centers last longer, but they're optimized for specific hardware generations. The entire build out needs to generate returns within five to seven years before it's surpassed by newer technology.

When the Music Stops

So what happens if the LLM scaling thesis breaks?

The first signs would be subtle. Model improvements from GPT-4 to GPT-5 are smaller than expected. Enterprise customers report that AI productivity gains have plateaued. The cost per incremental improvement keeps rising. OpenAI's revenue growth decelerates. Investors start questioning valuations.

Then the capital allocation committees at Microsoft, Google, and Meta face a decision. Do we commit the next $50 billion to AI infrastructure buildout? The 2024 and 2025 spending is already committed, facilities under construction, chips on order. But the 2026 and 2027 budgets come up for review in mid 2025.

If the return on AI infrastructure spending is unclear, CFOs will push to moderate. Not eliminate it, AI is genuinely useful, but reduce it from $50 billion annually to maybe $20 to 30 billion. Focus on maintaining existing infrastructure rather than expanding. That's a 40 to 50 percent reduction in incremental capex.

For Nvidia, this is catastrophic. Their data center revenue has grown from roughly $15 billion in 2022 to over $50 billion in 2024, driven almost entirely by AI accelerator sales to hyperscalers. If those customers cut spending by half, Nvidia faces an immediate demand collapse. Their stock, currently priced for continued exponential growth, would correct sharply.

The supply chain amplifies this. TSMC has invested billions in advanced packaging capacity specifically for AI chips. Memory manufacturers have ramped HBM production. Networking equipment vendors have tooled up for AI infrastructure. When demand drops, all of these see simultaneous revenue hits.

But here's what's interesting: the infrastructure doesn't disappear. It sits there, already built, already paid for, depreciating whether it's used or not.

The Stranded Asset Problem

When you build specialized infrastructure and demand evaporates, you have stranded assets. Oil refineries optimized for Venezuelan heavy crude when Venezuelan production collapsed. Fiber optic networks laid during the dot com boom. Shipping capacity built anticipating continued Chinese export growth.

AI training infrastructure could face the same fate. Massive data centers filled with GPUs, optimized for training large language models, suddenly underutilized because the AI labs don't need more training capacity and enterprise customers aren't scaling as fast as projected.

The economics change dramatically. If you built a data center expecting to run it at 80 percent utilization generating X revenue, but actual utilization is 40 percent, you need to find uses for the spare capacity. The capital cost is sunk. Now you're optimizing for operational cost recovery.

This is where materials discovery becomes interesting. The compute requirements are large but different from LLM training. You're not training monolithic models on exabyte scale datasets. You're running thousands of parallel simulations, ensemble calculations, multi scale models. The workloads are more varied, more intermittent, but still massively parallel.

For a hyperscaler with excess GPU capacity, materials discovery workloads become attractive. The marginal cost of running simulations on already installed hardware is just electricity and cooling. If you can sell that capacity at any price above marginal cost, it contributes to operational cost recovery.

Here's the scenario: it's 2027. The LLM boom has cooled. Microsoft, Google, and Amazon have data centers running at 50 percent utilization. Materials discovery startups and corporate R&D labs are willing to buy compute time for DFT calculations, molecular dynamics, inverse design searches. The hyperscalers start offering discounted rates for these workloads, potentially 60 to 70 percent below peak LLM training prices.

Suddenly, approaches that were economically infeasible become viable. Running ensemble DFT with 100 different functionals per material? Too expensive at $1 per GPU hour, but interesting at $0.30. Systematically mapping millions of material compositions? Not worth it when you're competing with OpenAI for capacity, but compelling when there's slack.

The materials discovery companies that raised at inflated valuations during the AI boom now have access to cheap compute. They couldn't justify the spending at 2024 prices, but 2027 prices change the economics. Projects that needed $100 million in compute budget now need $30 to 40 million.

The Economic Transition

This creates an interesting dynamic. The LLM hype cycle drives massive infrastructure buildout from 2023 to 2026. That buildout overshoots actual LLM demand by 2027. Prices for compute time collapse as hyperscalers try to monetize stranded assets. Materials discovery, which has genuine compute needs but couldn't afford peak pricing, suddenly becomes economically viable at scale.

It's reminiscent of the dot com boom and bust. Massive overinvestment in fiber optic infrastructure during the late 1990s. Crash in 2000 to 2002. But all that dark fiber still existed. Bandwidth prices collapsed. Suddenly applications that weren't economical at peak pricing became viable. YouTube wouldn't have been possible without cheap bandwidth, which only existed because of the prior overinvestment.

The same pattern could play out with AI compute. Overinvestment chasing LLMs creates excess capacity. Prices fall. Applications with genuine but less hyped use cases, like materials discovery, inherit the infrastructure at a fraction of the original cost.

The economics of materials R&D shift fundamentally. Instead of spending $50 million on experimental synthesis and characterization, you spend $30 million on computation to narrow the candidates, then $20 million on targeted experiments. The total cost is the same, but you're testing far more candidates computationally before committing to expensive experimental validation.

This doesn't solve the fundamental problems I outlined in my technical post. DFT is still inaccurate for complex systems. Models trained on small molecules still don't generalize perfectly to large alloys. Crystal structure prediction remains difficult. But with compute costs dropping 60 to 70 percent, you can afford to be less efficient. Run more calculations, try more approaches, validate more candidates.

Timeline and Catalysts

When does this transition happen? I think about it in funding cycles and budget planning timelines.

The current LLM infrastructure buildout is committed through 2025. Microsoft, Google, and Meta have announced capex plans, signed contracts with Nvidia, started data center construction. That spending happens regardless of whether AI delivers expected returns.

The critical decision point is late 2025, when 2027 budgets get set. By then, we'll have 18 to 24 months of data on enterprise AI adoption, revenue from AI products, productivity improvements from AI tools. If those numbers are strong, spending continues. If they're weak or plateau, spending moderates.

I'd estimate a 60 to 70 percent chance that AI infrastructure capex moderates significantly in 2027. Not because AI is useless, but because the returns on incremental spending diminish. Going from zero AI to current capabilities generated massive value. Going from current capabilities to 2x better capabilities might generate much less incremental value, while costing just as much.

When that moderation happens, you see a 12 to 18 month lag before it shows up in utilization rates. The 2026 buildout completes in 2027, coming online just as incremental demand growth slows. Suddenly you have new data centers ramping up while customer demand is flat or growing slowly.

That's when prices collapse. Hyperscalers start competing aggressively for workloads to fill capacity. Materials discovery companies, AI research labs doing non LLM work, anyone with parallel compute needs can negotiate steep discounts.

The transition from LLM infrastructure to materials discovery infrastructure happens gradually from 2027 to 2029. Not a sudden switch, but a slow repurposing as hyperscalers optimize for whatever workloads they can sell.

What This Means for Materials Startups

The timing implications for materials discovery startups are fascinating. Companies that raised in 2023 to 2024 at high valuations based on AI hype now need to survive until 2027 to 2028 when compute becomes cheap enough to make their approaches economical.

This is a classic bridge financing problem. You raised at a $500 million to $1 billion valuation expecting to demonstrate results quickly. But compute costs are too high to run the experiments you need. You're stuck doing smaller scale work, pilot projects, partnerships with industrial companies to access their computational resources.

Then in 2027, compute prices drop by 60 percent. Suddenly your runway extends dramatically. The $50 million you raised can buy 2.5x as much computation. Projects that weren't feasible become possible.

But you need to survive the gap. That means either raising bridge rounds (difficult if you haven't shown major progress), finding revenue through consulting or tools sales (pivoting from your original thesis), or drastically reducing burn rate (slowing research progress).

The companies that navigate this successfully will be those that recognized the timeline mismatch early. They didn't promise revolutionary materials discoveries by 2026. They positioned themselves as building platforms and tools, generating revenue from software sales to corporate R&D labs, accumulating smaller datasets and validating their approaches at modest scale, waiting for the moment when compute becomes cheap enough to scale up.

The companies that fail will be those that believed their own hype. They promised materials breakthroughs on VC timelines, spent aggressively on compute at peak prices, burned through capital without generating defensible IP or revenue, and find themselves out of runway just as the economic environment would have shifted in their favor.

The Broader Pattern

This pattern isn't unique to AI and materials. It happens repeatedly in technology cycles. Massive investment in infrastructure chasing one application, overshoot and correction, then the infrastructure gets repurposed for different applications at much lower cost.

The dot com boom built fiber optic networks for B2B exchanges and enterprise portals. Those businesses failed, but the infrastructure enabled YouTube, Netflix, and cloud computing. The telecom investment was economically terrible for the original investors but created enormous value for society by making bandwidth cheap.

The fracking boom overinvested in oil and gas production chasing $100 per barrel oil. Prices crashed, but the infrastructure and expertise remained. Natural gas became so cheap it displaced coal for electricity generation, reducing carbon emissions faster than any policy could have achieved.

The current AI boom might follow the same arc. Economically questionable for the original investors, but creating infrastructure that enables applications we haven't imagined yet. Materials discovery is one possibility. Computational biology, climate modeling, protein folding, there are many domains that need massive parallel computation but can't justify it at peak pricing.

The Contrarian Bet

Here's what I'd do if I were deploying capital today with a 5 to 10 year horizon.

Don't invest in the hyperscalers or Nvidia at current valuations. They're priced for continued exponential growth in AI compute demand, which I think is unlikely. The correction when it comes will be sharp.

Don't invest in materials discovery startups at inflated 2024 valuations. They're pricing in success on unrealistic timelines with compute costs that make their economics challenging.

Instead, wait for the correction in 2026 to 2027. When compute prices have fallen and valuations have reset, that's when materials discovery becomes genuinely interesting. You can invest in companies with realistic business models, access to cheap compute, and timelines that align with the actual difficulty of materials R&D.

Or even better, invest in the picks and shovels for computational materials science at the bottom of the cycle. High throughput experimental equipment manufacturers. Specialized software for materials simulation. Databases and standards for materials data. These businesses benefit from increased computational materials work regardless of which specific startups succeed.

The irony is that the LLM hype cycle, even if it disappoints, might accidentally enable a genuine revolution in materials discovery. Not because AI fundamentally solves the physics problems, but because overspending on LLM infrastructure creates cheap compute capacity that materials discovery can inherit.

We'll have built a trillion dollars of infrastructure chasing better chatbots and ended up with the tools to discover better batteries, magnets, and superconductors. Sometimes the best outcomes come from the most indirect paths.