The Concorde burned one ton of fuel per passenger to cross the Atlantic. One hundred seats. Three and a half hours. Mach 2. The most advanced commercial aircraft ever built — and every engineer who saw it wanted to believe it was the future.
The 747 did the same crossing in seven hours. Four hundred seats. A quarter of the fuel per passenger. No afterburners. No sonic boom. No government subsidies keeping it alive.
By 2003, the Concorde was a museum exhibit. The 747 had reshaped civilization.
The fastest plane ever built didn’t lose to a faster plane. It lost to a slower one that understood economics.
I keep coming back to this story because I’m watching it replay — in real time — in AI infrastructure.
The Concorde Bet
In the 1960s, the pitch was irresistible. Supersonic travel would make conventional aviation look like horse-drawn carriages. The British and French governments poured billions into a joint program. The engineering was extraordinary — delta wings, variable-geometry intakes, afterburning turbojets that pushed a narrow aluminum fuselage past the speed of sound at 60,000 feet.
It worked. Technically, it was a masterpiece.
Then the 1973 oil crisis hit. Fuel prices doubled. And suddenly, the math that was already unfavorable became brutal.
Concorde’s seat-mile cost was 90% higher than the 747’s. It carried 100 passengers; the 747 carried 400. Fuel consumption per passenger was four times worse — roughly 260 gallons per seat versus 65 on the 747 for the same Atlantic crossing. And because supersonic flight produces a sonic boom that rattles windows and terrifies livestock, Concorde was banned from overland routes. The only viable corridor was London–New York and Paris–New York. Two routes. That’s it.
The 747, meanwhile, could fly anywhere. It opened Tokyo, Sydney, São Paulo, Johannesburg. It didn’t just serve an existing market — it created one. Air travel stopped being a luxury and became infrastructure. The 747 democratized the sky.
Concorde carried diplomats and investment bankers willing to pay $6,000 one-way. The 747 carried everybody else — and changed the world while doing it.
The GPU Concorde
Now look at the AI infrastructure landscape.
The hyperscalers are spending $600 billion on capex in 2026 — Amazon alone committed $200 billion, Alphabet $175 billion, Meta $115 billion — and three-quarters of it is going directly into AI infrastructure: GPU clusters, networking fabric, power and cooling for facilities that consume more electricity than mid-sized cities.
The narrative is familiar: more compute, faster interconnect, denser racks, bigger training runs. Push the boundary. Chase Mach 2.
And the engineering is extraordinary. NVLink domains where 72 GPUs communicate at sub-microsecond latency over two miles of copper cable. Liquid cooling loops rejecting megawatts of heat from racks pulling 100 kW each. Power conditioning systems managing step-loads that would have tripped protection relays five years ago. I’ve written about this — the engineering is real, and it matters.
The raw numbers tell the story. NVIDIA’s data center GPUs have been doubling FP16 compute every 1.44 years — outpacing Moore’s Law’s traditional two-year cadence. From the P100’s 21 TFLOPS in 2016 to the B200’s 4,500 TFLOPS in 2024, the gap between actual GPU performance and what Moore’s Law would have predicted has blown open to roughly thirteen times.
Nobody disputes the trajectory. The Concorde’s engines were extraordinary too.
But engineering isn’t economics.
The question that killed the Concorde wasn’t “can we fly at Mach 2?” It was “can we fly at Mach 2 and fill enough seats at a price the market will pay?”
The question that will reshape AI infrastructure isn’t “can we build a 1 GW campus?” It’s “can we build a 1 GW campus and deliver tokens at a cost that sustains a business model?”
Concorde’s seat-mile cost was 90% higher than the 747’s. Right now, the cost of producing a token fell 280 times between 2022 and 2025 — roughly tenfold per year, faster than Moore’s Law, faster than solar panels, faster than any cost curve in industrial history. And yet total spending on AI infrastructure grew 320% over the same period.
Per-unit costs collapsed. Total expenditure exploded. Both things are true at the same time.
I wrote about this before. The pattern isn’t a contradiction. It’s a market finding its equilibrium. But equilibrium has a habit of arriving on its own schedule, and the companies still flying supersonic when it arrives will discover what British Airways discovered in 2003: the math doesn’t care about your engineering.
What Actually Won the Sky
The 747 wasn’t a safe bet either. Boeing nearly went bankrupt building it. The development program cost roughly $1 billion — around $7 billion in today’s dollars — and pushed Boeing’s total debt past $2 billion. A bet-the-company gamble. But the bet was on economics, not speed.
The 747’s design philosophy was the opposite of Concorde’s. Instead of pushing the envelope of what was physically possible, it asked: what does the market actually need? Volume. Move more people, more efficiently, over more routes. Make the seat cheap enough that a middle-class family could fly to Europe. The 747 didn’t try to eliminate the constraint of time. It collapsed the constraint of cost.
In my three-phase model of technology adoption, I argued that technology doesn’t win because it’s true — it wins when constraints fall. The Concorde pushed the performance constraint. The 747 pushed the economic constraint. Only one of those mattered at scale.
The parallel to AI infrastructure is direct. The companies building Concorde-class GPU campuses — maximum performance, maximum density, maximum wow factor — are making a real engineering bet. But the companies that figure out how to deliver inference at the lowest cost per useful token, reliably, across the widest addressable market — they’re building the 747.
And the 747 wins. It always wins.
TV Didn’t Kill the Radio
When television arrived in the 1950s, it looked like an extinction event for radio. TV stole radio’s advertisers, its celebrities, its prime-time programming, and its living room audience. Why would anyone listen when they could watch?
Radio should have died. It didn’t.
What happened was more interesting than survival. Radio found its layer. The transistor — invented at Bell Labs in 1947 — made radios portable. By the late 1950s, Sony was selling pocket radios. Suddenly, radio wasn’t competing with television for the living room. It was in the car, on the beach, in the bedroom, in the factory. Places television couldn’t reach. And instead of trying to replicate what TV did better, radio reinvented its format — long-form drama disappeared, music moved in, the DJ became the product, FM delivered fidelity that AM never could, and by the 1970s the medium had fragmented into dozens of formats serving audiences that television was too broad to address.
I think about this when people talk about AI making everything that came before it obsolete. Traditional compute, conventional data centers, standard enterprise infrastructure — the assumption is that these are the AM radio of the technology world, waiting to be replaced by the AI television.
They’re not. They’ll find their layer. Not every problem needs a 72-GPU NVLink domain. Not every question needs a trillion-parameter model. Not every business can justify $30 per GPU-hour to answer a query that a well-indexed database could handle in microseconds. Conventional infrastructure will find the workloads where cost-per-useful-work beats raw capability — the same way radio found portability and format specificity when the living room was lost.
The Equilibrium Nobody Can See Yet
Here’s where I stop pretending I have all the answers.
Hyperscalers are spending at historically unthinkable levels — 45 to 57 percent of revenue going to capex. The combined revenues of the largest AI-native companies (OpenAI at $20 billion ARR, Anthropic at $14 billion and growing over 10x annually) are still a fraction of the infrastructure being deployed to serve them. The market is supply-constrained, not demand-constrained — which means the bet is that demand will materialize at a scale that justifies the investment.
That’s a bet. Not a certainty.
And the assumptions embedded in that bet are worth examining under a light.
CoreWeave — the neo-cloud provider that went public in 2025 and reached a stunning $60 billion-range valuation — depreciates its NVIDIA GPUs over six years. Six years. Their competitor Nebius uses four. Amazon recently shortened the useful life of its own AI servers. Now, to be fair, depreciation schedules aren’t arbitrary — there are real engineering arguments for longer useful lives, and CoreWeave isn’t stupid. But analysts flagged their schedule as aggressive accounting that suppresses operating expenses and inflates the book value of GPU collateral backing billions in debt. The SEC filing makes for interesting reading.
Think about what a six-year depreciation schedule actually means. An H100 that shipped in late 2022 at roughly $30,000 per unit is assumed, on CoreWeave’s books, to retain economic value through 2028. But we just saw the chart — GPU compute is doubling every 1.44 years. By 2028, the H100 will be three doublings behind the frontier. That’s an eight-times performance gap. The Concorde didn’t face an eightfold efficiency disadvantage overnight. The H100 will.
And it’s not just the hardware curve. The cost of what that hardware produces is collapsing even faster. When GPT-4 launched in March 2023, OpenAI charged $36 per million tokens. By mid-2024, GPT-4o delivered comparable quality at $4 per million — a ninefold drop in seventeen months. DeepSeek undercut everyone further at $0.28 per million input tokens. Models that were frontier twelve months ago become commodity infrastructure, and the tokens they produce get repriced accordingly.
This is the dynamic that makes long depreciation schedules so fragile. It’s not that the hardware breaks — it’s that the economic value of what the hardware does erodes faster than the accounting model assumes. You’re depreciating the asset over six years, but the market is repricing its output every six months.
Concorde’s backers made the same bet. The demand for supersonic travel existed — investment bankers and diplomats genuinely wanted to cross the Atlantic in three and a half hours. But “demand exists” and “demand exists at the price required to sustain the infrastructure” are two very different statements.
I’ve spent twenty years watching infrastructure investments play out, and the pattern is remarkably consistent even when the technology changes and the scale shifts and the people in the room swear this time is different: the thing that looks revolutionary in the prototype phase gets repriced by economics in the deployment phase. The winners aren’t the ones who built the fastest system. They’re the ones who built the system that could sustain itself — economically, operationally, and thermally — at whatever scale the market actually demanded.
What the Builders Should Watch
I’m not predicting a crash. I’m not calling this a bubble. The AI infrastructure being built right now is real, and the capabilities it enables are genuine. But I’ve been in enough rooms where the engineering was flawless and the economics were fatal to know that one doesn’t guarantee the other.
If you’re building or operating AI infrastructure, the Concorde lesson distills to three questions:
What’s your cost per useful unit of work — and is it falling fast enough? Not cost per GPU-hour. Not cost per token. Cost per useful outcome for your customer. The 747 won because cost per passenger-mile was the metric that mattered. Find your equivalent.
Are you building for two routes or two hundred? Concorde could only fly London–New York. The 747 could fly anywhere. If your infrastructure only makes economic sense for the largest training runs, you’ve built a Concorde. The market will be won by infrastructure that serves the long tail — inference at scale, fine-tuning for enterprises, hybrid workloads that mix GPU and CPU efficiently.
Who pays when the economics shift? Fuel prices doubled after 1973 and Concorde never recovered. Energy costs, GPU depreciation, cooling requirements — these are variables, not constants. The infrastructure that survives isn’t the one optimized for today’s economics. It’s the one that has margin when the variables move. That’s not a prediction. That’s just how depreciation works when the underlying technology won’t sit still.
The Concorde was a triumph of engineering and a failure of economics. The 747 was a triumph of economics that happened to be good enough engineering. In the long run, “good enough engineering at the right price” beats “extraordinary engineering at the wrong price.”
Every time.