Goldman Sachs dropped a line this week that should have stopped every trading desk cold: AI model costs fell as fast in three years as PC prices did in fifteen.
Three years. Fifteen. That's not a trend line. That's a tectonic shift wearing a spreadsheet costume. I don't care how many times you've heard "AI is getting cheaper" โ you haven't heard it framed like this. And the framing is exactly where the trap lives.
Because here's what the headline wants you to feel: margin collapse. Tech giants squeezed. The trillion-dollar infrastructure bet suddenly wobbling. Every crypto AI token that rode the "compute is scarce" narrative looks naked by Monday open.
Here's the problem. The number itself โ the raw Goldman data โ is neutral. It's a cost curve. It doesn't have an opinion. The opinion got bolted on afterward, by aggregators, by crypto news wires hungry for a bearish hook, by everyone who needed a villain for their AI-bubble thesis.
The 2017 break didn't teach me to trust the first headline. It taught me to read the second paragraph, and then go find the report nobody linked.
Goldman's framing is a borrowed authority play. The bank published a cost trend; the wire services condensed it into "costs are collapsing, profits are squeezed." Two sentences, one direction. The bull case โ that cheap intelligence explodes total demand โ got cut for length.
And the source matters. This landed on a crypto news desk, not a semiconductor research desk. That means the AI content is aggregated, re-translated, stripped of its original boundaries. When you compress a cost curve into a headline, you lose the one thing that makes it useful: the definition of what's being measured.
So let me reconstruct what's actually being claimed. Between 2021 and 2023, the inference cost of GPT-3-class capability fell somewhere between 50x and 280x per year, depending on whose benchmark you trust โ a16z ran the early numbers, Stanford's AI Index confirmed the direction. A three-year window delivering a fifteen-year PC-style decline is aggressive but not fiction. The direction is real.
What's not real is treating "AI cost" as one number. There are at least three things hiding under that label, and they move in opposite directions.
The PC analogy is the tell. PC prices fell for decades โ and the PC industry became a multi-trillion-dollar market. Falling prices and exploding markets are not enemies. They're usually the same story. Anyone citing the PC parallel to argue "cheap AI kills the AI trade" is quoting half the sentence.
Now โ why now? Because the capex is at the point where it has to be justified. North American hyperscalers committed roughly $200 billion to AI infrastructure across 2024 and 2025. That money needs a return. A cost-collapse headline is the perfect stress test: either cheap intelligence expands the pie, or it shrinks the revenue pool and the depreciation eats everyone.
I've watched this movie before. In 2020, during the DeFi summer, I built a crude Python script to watch Uniswap V2 reserve changes in real time โ not because I trusted the model, but because I wanted to feel the market's pulse before the charts confirmed it. The lesson stuck: cost curves and price curves are different animals, and the gap between them is where traders get slaughtered.
Start with the four engines behind the decline. They stack, and none of them is a single breakthrough.
First, hardware generations. H100 to H200 to B200 to GB200 โ every hop improves compute per dollar and, critically, energy efficiency and interconnect bandwidth. That's the floor of the curve.
Second, architecture. Mixture-of-Experts shrinks activated parameters per token. Grouped-query and multi-query attention compress the KV cache. Attention variants keep trimming the quadratic tax. These are structural wins, not tuning.
Third, inference engineering โ this is where the real bleeding happens. FP8, INT4, AWQ quantization. Speculative decoding. Continuous batching. Prefix caching. FlashAttention kernels. Together they cut the compute and memory overhead per token by multiples. If you've ever wondered why your API bill dropped without the model getting smarter, this is why.
Fourth, distillation. Seven- and eight-billion-parameter models now match last-generation giants on most tasks. Capability that used to require a rack now runs on a single card.
The headline conflates three costs that don't move together: inference cost, training cost, and cost-per-capability. Inference is falling fast. Training is going the other way โ frontier training compute is still climbing exponentially. And cost-per-fixed-capability is collapsing while cost-for-the-frontier is not. Goldman's curve almost certainly measures the first. The story everyone tells measures the third.
That distinction is not academic. It's the whole trade.
Here's the second layer nobody prices: price is not cost. API prices for GPT-4-class capability fell from roughly $30 per million tokens to single digits โ sometimes lower โ across 2023 and 2024. Claude, Gemini, DeepSeek, the Llama family all followed. But some of that decline is competitive subsidy, not efficiency. Vendors eating margin to hold share. When the gap between price and cost is funded by a balance sheet, it is not a durable trend. It's a marketing budget.
And the open-weight pressure is the commercial engine underneath all of it. Llama, Mistral, Qwen, DeepSeek give away capability that used to be a moat. You cannot charge rent on something that's free. That's the real reason the model layer feels the squeeze โ not the cost curve, the commoditization.
The Chinese cost-efficiency shock deserves its own line. DeepSeek's high-value, low-cost route didn't just win benchmarks โ it reset the global pricing expectation. When a model trained on a fraction of the budget matches frontier performance, every closed API's price ceiling drops. That's not a technology story. It's a margin story.
Which brings us to the fight the headline skips entirely: training versus inference. Inference is deflating. Training is inflating. Those two facts point infrastructure investment in opposite directions. A single "costs are falling" line erases the tension.
Now the part that actually decides whether this is a boom or a bubble: Jevons. Cheaper intelligence should expand total consumption โ if demand elasticity is greater than one. If it is, falling unit costs drive rising total compute spend. Data centers, power, networking all get pulled harder. If elasticity is below one, cheaper AI just shrinks the revenue pool while the depreciation clock runs.
Nobody has settled this. The bulls quote Jevons. The bears quote margin compression. Both cite the same Goldman number. That should tell you the number is not the argument.
Ask the uncomfortable question: what are the unit economics right now? Is the model layer running positive gross margin or strategic loss? Nobody on the outside knows, and the companies have every incentive to blur it. If demand elasticity exceeds one, the total profit pool grows even as prices fall. If it doesn't, you're watching a race to zero with a capex bill attached.
On the model layer specifically, the value is migrating. When capability becomes a commodity, the capture point moves up the stack โ to applications, to agents, to proprietary data, to distribution. This is the operating-system playbook. Windows didn't win by being the best OS. It won by owning the distribution layer. Same movie, faster cuts.
For infrastructure, the nuance is that unit-token cost falling does not mean total compute spend falling. If anything, the cheaper the token, the more tokens get burned. The hard constraint shifts from silicon to electrons โ power and energy may bind before chips do. In most regions, data centers are already power-constrained before they're chip-constrained. If Jevons holds, the trade pulls toward megawatts, grid, and cooling as much as toward GPUs. That's why in-house silicon โ ASICs and domain-specific accelerators โ and inference-optimized hardware become strategic, not optional.
And the valuation transmission runs three ways. Application-layer businesses that treat AI as a cost line see margin expansion โ directly bullish. The model layer sees pricing power erode โ neutral to bearish. The compute supply chain depends entirely on whether Jevons holds. One data point, three opposite readings.
Here's the angle I don't see anywhere in the coverage.
Everyone is debating whether cheap AI kills margins or expands markets. The real bet underneath is subtler: the model layer may enter a value vacuum โ a window where commoditization moves faster than value can migrate up to the application layer. Not a crash. A gap. A stretch where the old rent is gone and the new rent hasn't been built yet.
That gap is where the pain lives. Not in hyperscalers โ they cross-subsidize AI against cloud revenue, so their AI pricing doesn't have to stand alone. Not in open-source โ they have no revenue to lose. The squeeze lands on independent model vendors caught between falling prices and rising training costs, with no distribution to fall back on.
And there's a deeper blind spot: the whole "cost collapse" narrative assumes efficiency drove the decline. But if a meaningful slice of the price drop is subsidy, then the demand expansion it supposedly fuels is built on sand. Pull the subsidy and the elasticity test never actually happened. You'd be measuring a promotional price, not a market-clearing one.
That's the risk nobody models. A cost curve driven by efficiency is a bull signal. A price curve driven by competition is a bear signal. The headline can't tell them apart โ and neither can most of the people trading on it.
One more thing the coverage skips: the efficiency gains are not evenly distributed. They're concentrated in inference, in specific task classes, in specific model sizes. The frontier โ the thing that wins the next benchmark โ is still brutally expensive. Anyone extrapolating the inference curve onto the frontier is drawing a straight line through two different datasets.
So watch the split, not the headline. Track hyperscaler capex against disclosed AI revenue โ that ratio is the ROI verdict. Track API pricing against reported margins โ the difference is who's paying for the subsidy. And track total inference volume and power draw โ that's the only real test of whether Jevons is winning.
The cost curve is real. The interpretation is still unclaimed. And the people who move fastest on the wrong reading are the ones left holding the depreciation.