We've seen this story before — and it didn't end the way anyone expected.
Somewhere right now, a construction crew is breaking ground on another data center. Maybe in Virginia. Maybe in Pune, or a corridor outside Jakarta chosen for its land cost and proximity to a power substation. The building will cost hundreds of millions of dollars. It will draw enough electricity to power a mid-sized town. It will be full of chips — expensive, power-hungry, imported chips — and it will be built on an assumption so widely shared that almost nobody has bothered to say it out loud.
The assumption is this: AI inference is a centralised business. You send a query to the cloud. The cloud thinks. The cloud answers. For the foreseeable future.
That assumption deserves a harder look than it's currently getting.
We Have Seen This Before
In the late 1990s, telecoms companies laid enough fiber optic cable to circle the Earth tens of thousands of times. The investment thesis was simple: the internet was growing exponentially, bandwidth was the constraint, and whoever owned the pipes would own the future. Companies borrowed billions. WorldCom became one of the largest corporations in America. Transoceanic cables were sunk across the Atlantic and Pacific. Entire cities were rewired underground.
Then the crash came. WorldCom filed for bankruptcy in 2002 — at the time, the largest in American history. Hundreds of smaller telecoms disappeared. The fiber glut was so severe that most of the cable laid never carried a single commercial byte. It sat dark, underground, forgotten.
Except it wasn't forgotten. Google found it. Netflix found it. A decade after the crash, those same fiber lines became the physical substrate of the streaming era, the cloud era, the mobile internet era. The overbuilders lost everything. The infrastructure they left behind changed the world.
The economist Carlota Perez has a framework for understanding exactly why this happens, and why it keeps happening.¹ In her analysis of successive technological revolutions — railways, electricity, automobiles, and now digital — she identifies a recurring two-phase pattern. The first phase, which she calls Installation, is driven by financial capital chasing new infrastructure before anyone fully understands how to use it. Bubbles form. Overbuilding happens. A crash follows. The second phase, Deployment, is when society actually harvests the gains — when the infrastructure built speculatively gets absorbed into the productive economy and productivity spreads broadly.
The dot-com crash, on this reading, wasn't a failure. It was the installation phase completing itself. The dark fiber wasn't waste. It was infrastructure waiting for its deployment phase.
Peter Schwartz and Peter Leyden made a related argument in The Long Boom,² published in 1999 at the peak of the bubble. Their claim was that digital technologies were uniquely capable of sustaining multi-decade growth — not because crashes wouldn't happen, but because the underlying infrastructure being built was genuinely transformational. They were early and wrong about the timing. They were arguably right about the direction.
The question worth asking now is: where are we in this cycle with AI infrastructure?
What Just Changed
A 25-person startup in Toronto called Taalas has built a chip that runs a state-of-the-art AI model at 17,000 tokens per second. That's roughly 70 times faster than what Nvidia's best data center card delivers on the same model. It draws 200 watts — less than a high-end gaming PC. No liquid cooling. No specialised facility. It can, in principle, sit in a server room in a hospital in Chennai, or a government ministry in Delhi, or a bank in Lagos.
The technical reason is genuinely new: instead of loading a model's parameters from memory and processing them — the way every AI chip before it has worked — Taalas encodes the model's weights directly into the transistors of the chip during fabrication. The model doesn't run on the chip. The model is the chip.
You don't need to understand the engineering to understand the implication. If inference can happen locally, at this speed, at this power envelope, then a meaningful portion of what data centers are being built to do can move to the edge. To your office. Your hospital. Your government ministry. Your phone, eventually.
The data centers don't disappear. But the specific case for building this many of them, this fast, just got more complicated.
The Scale of the Bet
To appreciate what's at stake, consider the numbers announced in just the last two years.
Microsoft committed $80 billion to data center construction in 2025 alone. Amazon, Google, and Meta have made comparable commitments. In India, the government's Digital India push and hyperscaler expansion plans have seen data center capacity in Mumbai, Chennai, and Hyderabad growing faster than almost anywhere outside the United States. Across Southeast Asia, the Gulf, and Europe, the same story is playing out. Total announced global investment in AI infrastructure runs well past $500 billion.
These aren't software purchases you can cancel with a quarter's notice. They're land acquisitions, decade-long power purchase agreements, grid interconnect queues, and poured concrete. The depreciation schedules run 20 to 30 years.
The entire investment thesis rests on one thing: that demand for centralised AI compute will grow fast enough, and stay centralised long enough, to justify the build.
Both halves of that bet deserve scrutiny.
The Demand Question
The bull case for data centers has two engines: training and inference. Training — teaching models — stays centralised. You need enormous, coordinated compute to train a frontier model. Nothing about local silicon changes that.
Inference is different. It's the actual use of AI — answering questions, processing documents, running applications. By volume, the vast majority of AI compute demand. And it's the part most exposed to what Taalas represents.
If a meaningful portion of inference migrates to local silicon, the utilisation math on clusters built for the end of the decade starts to strain. Two forces are pushing inference costs down simultaneously — local silicon architecture and supply chain normalisation, as chip shortages and HBM constraints moderate through 2026 and 2028. Both compress the premium that currently justifies data center economics. Local silicon just accelerates what was already in motion.
The Clients Who Were Never Fully There
Governments, hospitals, law firms, banks, defence contractors — the highest-value customers for AI — have been conspicuously slow to adopt cloud AI. Not technophobia. Sending sensitive queries to a server in another country, owned by a private company, subject to foreign law, is genuinely untenable for much of what these institutions do.
This isn't abstract in India, where data localisation debates have run for years and government agencies are acutely aware of sovereignty questions. It applies equally in Indonesia, Brazil, Saudi Arabia, and the EU — large markets that want AI but have structural reservations about routing sensitive workloads through foreign-owned infrastructure.
Sovereign cloud offerings from AWS and Azure partially address this, but still require institutional trust in a third party. Local inference silicon changes the calculus entirely. A chip in a hospital in Bengaluru processes patient data in Bengaluru. A government ministry in Delhi keeps its documents in Delhi. The operational model can still be service-like and subscription-based — the hardware is just local, dedicated, and auditable in a way a cloud API never fully is.
This market was always going to be hard for centralised cloud to capture. It now has a credible alternative.
The Bull Case Isn't Dead
It's worth being direct about what this argument is not.
Data centers are not going away. A large part of the world — including most of rural India, Southeast Asia, and sub-Saharan Africa — is not yet actively online. As billions more people and devices connect over the next decade, raw infrastructure demand grows regardless of what happens at the architectural frontier. The payoff periods on today's buildouts may simply be longer than investors currently model — not negative, just slower.
This is, in fact, how the fiber story resolved. The overbuilders went bankrupt. The infrastructure they built became the foundation for a generation of companies they never imagined. Perez's deployment phase arrived, just not for the people who funded the installation.
And plenty of inference stays centralised. Applications where model sizes exceed what local hardware can handle, where shared infrastructure economics still win, where users don't require privacy or low latency — all of that remains cloud territory for the foreseeable future.
The question isn't whether data centers have a future. They clearly do. The question is whether the specific assumptions driving the current pace of construction — assumptions about permanent centralisation, stable inference margins, and a captive high-value customer base — are as solid as the capital commitments they're underwriting.
What Needs a New Argument
The original case for this wave of buildouts went roughly: AI inference demand is growing exponentially, it's structurally centralised, and the high-value customers will eventually come around. That case made sense in 2023.
In 2026, one of those three legs is wobbling, one is softening on its own, and the third just got a credible structural alternative.
Perez would recognise the pattern. Schwartz and Leyden would probably say the long boom is still coming — just not through the mechanism currently being funded. Both might be right. The fiber is being laid. Whether the companies laying it survive to see the deployment phase is a different question entirely.
Markets often react before the physical reality catches up. The construction crews keep digging either way. But somewhere in a boardroom in Mumbai, or a capital allocation meeting in Singapore, or an infrastructure fund review in Riyadh, somebody should be asking whether the model they're building for still looks like the model that's arriving.
The buildout case hasn't collapsed. It just changed. And the new version hasn't been written yet.
PS: Co-written with Claude
Next: the chip that started this conversation — how a 25-person Toronto startup built the thing that changed the question.
Notes
¹ Carlota Perez, Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages (Edward Elgar, 2002). Perez's framework identifies five successive techno-economic paradigms in modern capitalism, each following a two-phase pattern of Installation (speculative bubble) and Deployment (productive diffusion).
² Peter Schwartz and Peter Leyden, "The Long Boom: A History of the Future, 1980–2020," Wired, July 1997; expanded as The Long Boom: A Vision for the Coming Age of Prosperity (Perseus Books, 1999). Written at the peak of the dot-com boom, the piece argued that digital technologies were capable of sustaining multi-decade global growth — a claim that proved directionally correct even as the timing and mechanism diverged significantly from their prediction.
Comments
Post a Comment