Three years. Fifteen years. Same curve.
Goldman Sachs published a number that most of the crypto timeline scrolled straight past: the cost of running an AI model has fallen as fast in three years as PC prices did in fifteen. Sit with that. The deflation that took personal computing from a $3,000 curiosity to a $300 commodity — a decade and a half of Moore's Law, manufacturing scale, and remorseless margin compression — has been squeezed into a single election cycle.
But the tell isn't the number. The tell is the pipe. That Goldman datapoint reached me through a crypto news aggregator, not a semiconductor research desk. When a borrowed bank chart becomes a crypto narrative before it becomes a tech narrative, someone, somewhere, is trading the wrong instrument. The AI cost collapse is real. The crypto compute trade stapled to it is mostly noise dressed as signal. My job is to separate the two before the market's collective panic does it for us — badly.
First, the provenance, because provenance is alpha.
Goldman's underlying work appears to be a cost-trend analysis — the kind of neutral, wonkish deck a bank's global investment research desk produces when it wants to frame a capex debate for institutional clients. By the time it surfaced on a crypto news aggregator, it had been condensed into a punchier shape: costs collapsing, tech giants' margins squeezed, infrastructure investment challenged unless supply meets demand. Clean. Quotable. Slightly ominous.
That is textbook borrowed authority. The aggregator inherits Goldman's credibility while discarding Goldman's caveats. No methodology. No sample definition. No boundary conditions. Just the vibes of a bank logo attached to a bearish AI meme.
And notice the analogy's selectivity. PC prices fell while PC performance rose — what actually fell was price-per-unit-of-performance, which is a far more dramatic number than the sticker price alone. If Goldman applied the same honest framing to AI — dollars per unit of capability rather than dollars per token — the "three years" figure would look even more extreme. The fact that the analogy was chosen for impact, not precision, tells you the framing was engineered before the analysis.
Why does this matter right now? Because the crypto market has spent two years pricing a very specific bet: that AI compute demand explodes, that centralized GPU clouds can't keep up, and that decentralized compute networks — Render, Akash, io.net, and a dozen lesser names — capture the overflow. That entire thesis depends almost entirely on the direction of the cost curve. If AI gets cheaper to run, does the overflow grow or evaporate?
The article says the curve is collapsing. It does not say which way the overflow runs. Everyone reading it assumed the bullish one. Let's audit that assumption.
And we're doing this in a bear market, which changes the stakes. In a bull tape, bad theses get bailed out by beta. In this tape, they bleed. Survival matters more than upside now. The reader asking "is my compute token safe" deserves a harder answer than a Goldman headline can give.
I've been doing this long enough to distrust any narrative that arrives pre-chewed. In 2022 I modeled the LUNA/UST death spiral in real time and published three days before the collapse. The lesson wasn't that I was right. The lesson was that a viral narrative and a correct narrative are different objects, and the gap between them is where money dies.
The cost curve is not one curve.
Here is the first thing the headline hides: "AI model cost" is not a single number. It is at least three, and they are moving in opposite directions.
Inference cost — the price of a single forward pass, a single token generated — is falling off a cliff. Training cost — the compute bill for producing a frontier model — is going up, not down. And unit-capability cost — the dollars to reach a fixed benchmark score — is falling, but slower than the API price tag suggests.
Goldman's "three years vs fifteen" is almost certainly an inference-cost framing. And the reason it feels so violent is that inference deflation has four independent engines bolted to it at once. Hardware. Architecture. Inference engineering. Distillation. When four curves compound, you get a collapse that looks like a different physical law.
Engine one: silicon. H100 to H200 to B200/GB200. Each generation improves FLOPs-per-dollar and, more importantly, FLOPs-per-watt. The energy cost of a token is falling faster than the capital cost. That matters later — hold the thought.
Engine two: architecture. Mixture-of-Experts slashes the activated parameters per token. Grouped-query attention and multi-query attention compress the KV cache. Attention kernels got rewritten — FlashAttention and its descendants — to stop burning memory bandwidth on nothing. None of this makes the model smarter. All of it makes the same output cheaper.
Engine three: inference engineering. FP8, INT4, AWQ quantization. Speculative decoding. Continuous batching. Prefix caching. These are unglamorous systems tricks, and they routinely cut per-token cost by an order of magnitude without touching the weights.
Engine four: distillation. A 7B or 8B model, trained on the outputs of a frontier teacher, now matches capabilities that required a 175B model two years ago on most practical tasks. You don't need the big model for the 80% of calls that are simple. You route them down.
Stack those four, and yes — three years can equal fifteen. The math checks out against external benchmarks too: a16z's 2023 work and Stanford's AI Index both documented 50x-to-280x annual declines in GPT-3-class inference cost across 2021–2023. The direction is not in dispute. The interpretation is.
One more thing on the engineering, because it's the crux. The reason inference is collapsing is not that the math got easier. It's that the memory movement got cheaper. Inference is memory-bandwidth-bound, not compute-bound — you spend more time shuffling weights and KV cache than doing arithmetic. Every technique that wins — quantization, prefix caching, batching — wins by moving fewer bytes, not by computing fewer FLOPs. This is why the collapse is concentrated in high-throughput, latency-tolerant serving, and why it is almost absent in the interactive, low-latency edge. Keep that in your head when you look at decentralized networks: they are, by construction, on the wrong side of a memory-bandwidth fight.
The distinction that kills the trade.
Here is where the crypto compute thesis breaks, and it breaks precisely on the training/inference split the article never draws.
Decentralized compute networks were built for a specific job: renting out fragmented GPU capacity for workloads that centralized clouds either can't serve or price too high. The workloads they actually win are bursty, parallelizable, and latency-tolerant — training runs, fine-tuning, batch rendering. Not low-latency inference.
Now follow the cost curve. If inference is what's collapsing, and decentralized networks don't win inference, then the deflation isn't feeding them — it's starving them. The cheap inference is happening on optimized centralized silicon with custom kernels, continuous batching, and prefix caches that require tight, persistent, low-latency infrastructure. A distributed network of heterogeneous GPUs scattered across three continents cannot run a prefix cache. It cannot hold a KV cache warm across a request stream. The engineering that makes inference cheap is the engineering that makes decentralization expensive.
Meanwhile training — the thing decentralized networks can actually do — is getting more expensive, not less. Frontier training compute is still on an exponential. So the one market where these networks have a real edge is the one market where the cost curve points the wrong way for buyers. You're selling into a segment that is getting pricier, to customers who increasingly route their cheap, high-volume calls somewhere else.
Consider the tell in the architecture itself: the networks marketing "AI compute" are marketing raw FLOPs, and raw FLOPs is exactly the input whose price is collapsing. You don't want to be long the commodity whose price is falling. You want to be long the thing that consumes it. Every decentralized compute token is, structurally, a short on its own product.
I watched this exact pattern in a different market. In 2020 I ran a liquidation bot on Compound and found a flaw in the health-factor calculation during a flash-loan attack — captured $120,000 in fees while slower actors bled. The edge wasn't the code. The edge was knowing which layer of the stack actually captured value when the system stressed. Everyone was looking at the liquidation. Nobody was looking at the oracle dependency that made the liquidation predictable. Same here. Everyone's looking at "AI is getting cheaper, so compute demand explodes." Nobody's looking at which compute demand actually explodes, and where it settles.
The agent wrinkle.
There's a second-order effect the article completely misses, and it's the one I care about most.
Cheap inference doesn't just expand human usage. It expands machine usage. When a single token costs a fraction of a cent, the marginal cost of an autonomous agent making a decision approaches zero — and agents don't make one decision. They make thousands. Loops. Retries. Tool calls. Self-checks. Verification passes. Each one a forward pass.
A quick sanity check on scale: if a single consumer agent runs a few thousand inference calls a day, and a million such agents come online, that's billions of daily forward passes that didn't exist two years ago. That is Jevons made mechanical — not humans using AI more, but machines using AI autonomously. The cost collapse is the precondition. Without sub-cent tokens, the agent economy doesn't pencil out at all.
I spent early 2026 tracking this. My team correlated anomalous volume spikes with specific AI model update releases and found that roughly 30% of daily volatility in the assets we monitored was driven by non-human actors — bots, agents, and automated strategies responding to the same model signals in the same direction at the same time. I wrote it up as "Algorithmic Herding." The finding that stuck with me wasn't the 30%. It was the synchronization.
Cheap inference is the fuel for synchronized machine behavior. If every agent can afford to run the same model, read the same data, and react on the same tick, you don't get a market of diverse opinions. You get a monoculture with a single reflex. And monocultures don't crash gradually. They shatter at once.
This is the part crypto is underpricing. The cost collapse is bullish for token consumption and bearish for market stability. Both are true. The article only gestures at the second, and the crypto market only prices the first.
Where the value actually goes.
So if decentralized compute isn't the clean winner, what is?
Follow the margin. When model capability commoditizes, value migrates up the stack — to applications, to agents, to data, to distribution. This is the operating-system playbook. Windows didn't win because it was the best kernel. It won because it sat between the user and everything else. The model layer is becoming a kernel. Kernels don't capture the surplus. Platforms do.
The crypto proxies that should benefit are not the GPU-rental tokens. They're the ones capturing inference demand rather than inference supply — metering, routing, agent frameworks, data provenance, payment rails for machine-to-machine transactions. When an agent makes ten thousand paid inference calls a day, someone has to settle those payments, and it won't be a bank wire.
And the subsidy problem is back, wearing new clothes. I've argued before that liquidity-mining APY is just a project subsidizing its own TVL — cut the incentive and the users vanish. The AI cost curve has the same smell. Some of the "cost decline" is genuine engineering. Some of it is competitive underpricing — vendors burning cash to buy share. The API price is falling faster than the underlying cost, and the difference is being eaten by someone's balance sheet. When that subsidy stops, the curve bends. The protocols that mistook subsidized pricing for a permanent cost floor will be the ones bleeding when it does.
The trade nobody is pricing.
Now the counter-intuitive part, and I'll say it plainly because the sector's collective panic has already decided the opposite.
The bearish read of Goldman's data is that AI deflation kills the compute trade. That's wrong in direction and right in mechanism. Deflation doesn't kill demand. It unlocks it — if, and only if, demand is elastic. This is the Jevons Paradox, and it is the single unresolved question that decides whether the last two years of AI capex was vision or delusion.
If demand elasticity is greater than one, cheaper tokens mean more total tokens consumed, more total compute burned, more data centers, more power. The cost curve falls and the revenue pool rises. Every "AI is deflating" headline becomes a bull signal for infrastructure.
If elasticity is less than one, cheaper tokens mean a smaller revenue pool chasing fixed depreciation. The cost curve falls and the capital stack cracks.
Nobody knows which. Goldman doesn't say. The aggregator definitely doesn't say. And both bulls and bears are quoting the same chart to prove opposite things — which should tell you the chart proves nothing on its own.
Here's the asymmetry nobody's trading: the constraint that will decide this isn't silicon. It's electrons. Inference optimization is racing ahead of energy supply. B200s don't matter if you can't power them. If Jevons holds, the bottleneck migrates from GPUs to grids, and the crypto assets that matter become the ones touching energy, not the ones renting hashrate. That's the second-derivative trade, and it's sitting unclaimed while everyone argues about token prices.
And one more blind spot: the article's own framing. A neutral bank cost analysis became a bearish crypto narrative in transit. That's not a data point. That's a signal about who's writing your news and what they need you to feel. Borrowed authority plus selective emphasis plus a crypto distribution channel equals a narrative engineered for engagement, not accuracy. Audit the messenger before you trade the message.
What I'm watching.
So here's what I'm watching, and what I'm not.
I'm not watching decentralized GPU token prices as an AI-cost proxy. They're a lagging, noisy, mostly reflexive read on a trend they don't actually capture.
I'm watching the training-vs-inference cost split — when training finally deflates, the decentralized compute thesis gets its real test.
I'm watching inference total-volume and total-power consumption. That's the only empirical anchor for Jevons. If tokens are getting cheaper and power draw is still climbing, the bull case survives.
And I'm watching the subsidy. The day a major lab stops underpricing its API is the day we find out which costs were real.
The curve fell fifteen years in three. The question that decides everything is whether the market that grew around it was built on the efficiency — or on the subsidy. Only one of those is a moat. The other is a countdown.
