Here is the reality: Goldman Sachs published a data point claiming AI model costs fell as fast in three years as PC prices did in fifteen. A headline like that travels through the crypto desk the way a rumor travels through a validator set โ fast, loud, and rarely checked against the underlying state. I check. Not because the number is wrong; the number is directionally correct. But a cost curve without a schema is not a fact. It is a headline with a heartbeat.
Auditing isn't about finding intent. It's about tracing the cost curve to its mechanical source and asking who eats the difference between the posted price and the true cost.
The claim covers the period from roughly 2021 to 2024, when GPT-3-level capability moved from an expensive API call to something a startup could serve on a laptop. Third-party benchmarks confirm the direction. a16z's inference trackers and the Stanford AI Index both measured annual compounding declines in the 50-to-280x range for fixed-capability inference. The direction is real. The magnitude is real. What is not real is the clean story wrapped around those numbers: that AI is deflating the way PCs deflated, with a smooth slope and a guaranteed happy ending.
Context matters here, and my read on this curve comes from a specific history. In 2017, I spent nights in an Austin co-working space manually auditing Solidity source code for the first wave of ERC-20 launches. I found integer overflow flaws in three major token projects and collected $12,000 in bug bounties. That habit โ verify the surface metric against the mechanical structure underneath โ is how I have read every market story since. When a crypto outlet relayed Goldman's curve, I applied the same reflex. Which unit is the "cost" measured in? Dollars per million tokens? Dollars per training run? Dollars per unit of benchmark capability? The summary never says, because the fastest-falling number is the one quoted, and the fastest-falling number is rarely the number that matters.
The production function of the decline is a four-layer stack, and every layer has a blockchain analogue.
Layer one is hardware generation. H100 to H200 to B200: each generation cuts unit compute cost and improves energy efficiency per FLOP. This is the AI equivalent of ASIC generations in Bitcoin mining. When I audited a mining pool's cost model in 2018, the curve had the same shape: each node release roughly halved energy per hash, and miners who refused to upgrade were short volatility without knowing it. Hardware is the slowest driver but the most powerful, because it resets the baseline for every layer above it.
Layer two is architecture. Mixture-of-experts systems activate only a fraction of parameters per token. Grouped-query attention and multi-query attention compress the KV cache. Attention kernels were rewritten from first principles. This is not optimization; it is shape-shifting. The computation gets restructured, not merely sped up. The crypto equivalent is the jump from squeezing gas costs on a sidechain to rolling thousands of transactions into a single validity proof. One is incremental. The other changes the cost structure of an entire chain.
Layer three is inference engineering. FP8 and INT4 quantization, speculative sampling, continuous batching, prefix caching, a dozen kernel families that each shave percentages off serving cost. These are where an uncomfortable share of the observed decline actually lives. I learned this pattern in DeFi Summer 2020, when I deployed $50,000 into Uniswap V2 and Curve and spent weeks backtesting impermanent loss strategies with custom Python scripts. The discovery was unglamorous: most apparent efficiency in a protocol is not a breakthrough. It is someone finally removing the stupid bottleneck. A batch scheduler that processes 64 requests instead of four is not a research miracle. It is a plumbing fix with a nine-figure PR budget.
Layer four is distillation. Seven- and eight-billion-parameter models now approximate the capability of a frontier model from a generation ago on most practical benchmarks. Small models inherit large behaviors through distillation the way a Layer2 inherits security from its base layer. Their serving cost is an order of magnitude lower, and the capability gap keeps compressing. Every vendor is racing cheap inference to the edge, and the edge is where the volume will land.
There is also a fifth actor hiding in the shadows: the decentralized compute marketplaces. Every network that sells GPU time through a token โ the Render of the world, the Akash of the world, the compute subnets of Bittensor โ depends on the same cost curve. But token incentives distort the signal. A subsidized compute token is not the spot price of FLOPs; it is a marketing instrument wearing a futures contract. Every time the unit of account is an emitted token rather than a metered FLOP, the cost signal becomes a narrative. In a bull market, these networks price optimism. In a bear market, they price utilization. The cost curve is the same in both phases โ only the token multiple changes.
Now the metric problem the Goldman summary ignores: inference and training costs are not moving in the same direction.
Inference is deflating at an absurd rate. Frontier training is inflating at an equally absurd rate. The compute required to train the strongest model is still growing exponentially; a single frontier run now costs nine figures. The summary presents one smooth curve. The real system has two diverging curves. Reasoning is commoditizing. Frontier creation is becoming a superpower-level capital project. Those two statements imply opposite infrastructure positions. Conflating them produces losses in both directions.
This bifurcation has a direct blockchain analogue. Execution on a legacy Layer1 used to be the expensive bottleneck. Rollups collapsed execution cost by orders of magnitude while settlement and proof verification stayed expensive. The people who understood that split built the L2 books. The people who priced "blockchain cost" as a single number got run over by the fee markets. AI is running the same playbook at a different altitude: the execution side is a trillion-dollar deflation story, and the settlement side is the question nobody wants to name โ how do you verify what a model actually computed before you pay for it?
I have to flag my own bias here. ZK proving costs are still absurd. In a bull market, proving expense hides inside inflated token economics; in a down market, operators bleed. I have been saying this since the rollup wars: cheaper execution without cheaper verification just moves the bottleneck. The parallel to AI is uncomfortable. The cost of generating the result collapses. The cost of proving the result has not collapsed at the same rate. Verification is the uncheapened part of every cheapening stack.
Now pull the second thread: price is not cost, and the distance between them is where the real signal lives.
Between 2023 and 2024, the API price of GPT-4-class inference collapsed from roughly $30 per million tokens to single digits and below. A significant share of that decline is competitive discounting, not engineering. Closed-vendor pricing was disciplined by open-weight models โ Llama, Qwen, DeepSeek โ that offered credible free alternatives. Open weights function as the audit floor of the model market: they keep the closed labs honest the way an open-source client keeps a consensus network honest. The price curve fell faster than the cost curve, and the difference is a subsidy paid out of investor capital, cloud-bundle accounting, and strategic patience. A hyperscaler that wraps AI pricing into cloud commitments has already left the notion of marginal cost behind. At that point, the sticker price stops being a sound and becomes a sales instrument.
Crypto has seen this script. It is the liquidity mining playbook. In 2020, yield farms printed tokens to subsidize liquidity that evaporated the moment the incentive schedule leaked. The yield curve looked like an efficiency miracle until the subsidy decayed and the underlying pool re-priced. A token price subsidized by emissions is not an efficiency curve; it is a burn curve with better charts. An API price subsidized by cloud bundling is the same instrument wearing a different suit. Goldman's summary says infrastructure investment is being challenged. That phrasing is the anxiety showing through. The market knows there is a gap between advertised deflation and engineering reality, and the entire capex thesis now rests on whether that gap gets closed by genuine efficiency or blows up into a price war that no model vendor survives at scale.
Which is the entry point for Jevons paradox.
Cheaper AI increases total AI demand. If demand elasticity is above one, total compute consumption rises even as unit prices collapse, and the roughly two-hundred-billion-dollar annual hyperscale capex program turns out to be rational. The PC analogy works exactly this way: prices fell, the market exploded, and a trillion-dollar ecosystem emerged that no one predicted from the price curve alone. Look only at the price curve and you see margin compression. Look at the total compute spend curve and you see a demand volcano. The paradox is the difference between being long the compute chain and being short it.
But the paradox has a failure mode. If demand elasticity is below one, the deflation shrinks the total revenue pool and the capex gets stranded. And the uncomfortable truth is that nobody knows the elasticity number. Not Goldman. Not the crypto outlet quoting Goldman. Not the hyperscalers writing the checks. I spent the 2022 bear market tracing on-chain ledgers of failed lending protocols, mapping $2 billion in locked value back to centralized oracle manipulation rather than smart contract bugs. That experience built my respect for the gap between narrative and structure. Flow follows fear, but only if the protocol holds. AI demand follows cheap prices, but only if the infrastructure holds its margin. A demand boom built on subsidized prices and deferred depreciation can reverse faster than any confidence interval predicts.
There is a measurement problem buried in this too. Hyperscalers report AI capital expenditure quarterly, but they do not report the unit economics of their inference products. The ratio of AI revenue to capital expenditure โ the single number that would tell you whether the Jevons bet is working โ is disclosed nowhere. Nobody wants to publish that table. In crypto, we call that silence the loudest audit trail in the market. Silence is the loudest audit trail in the market.
Now the contrarian angle, because the crypto-native read diverges from the Wall Street read exactly at this point.
The blind spot is not the cost curve. The blind spot is the provenance curve. As inference gets cheaper, the cost of generating false data collapses at the same rate. Deepfakes, automated phishing, fabricated benchmarks, AI-written media engineered to move markets โ every product in that category rides the same deflation slope as legitimate inference. This is the security externality that every cost-deflation discussion ignores. The ledger doesn't lie, but the training set might. And the API response might. And the article quoting the API response might. In a world where synthesis is cheap, the scarce resource is no longer intelligence. It is verifiability.
This is where my current work points. I founded a community called Verifiable Truth, centered on zero-knowledge proofs that bind an AI output to the provenance of its training data. We built a prototype that lets a model attest to the origin of its training corpus without revealing the corpus itself. From where I sit, that is not a privacy feature. It is the settlement layer for the machine economy. If any agent can mint plausible text at near-zero cost, the only defensible scarcity is cryptographic attestation: proof of where the data came from, proof of what the model computed, proof that an output is not a hallucinated piece of market-moving garbage.
There is an institutional angle here worth naming. In 2025, I worked with a legal engineering team to draft a "Proof of Decentralization" standard for the Texas State Blockchain Council โ a framework quantifying node distribution and governance participation so regulators could verify decentralization instead of guessing at it. The lesson carried over directly. Every institutional adoption wave eventually demands a verification standard, and the standard only works if it is built into the protocol from the start. Enterprise AI adoption will hit the same wall. Firms will not trust model outputs because of a brand promise. They will trust outputs that carry a cryptographic receipt. That is the bridge between the cost collapse and the compliance era.
The uncomfortable conclusion for the model layer is that cost decline is commoditization, and commoditization transfers value.
The winners will not be the vendors printing the deflation curve. They will be the application layers, the agent frameworks, the data owners, and the verification infrastructure. I have watched this migration inside crypto itself. Base layer gas got cheap; value migrated to applications, aggregators, and middleware. The same sequence is running now. Three years of falling AI prices is not a signal that model vendors are about to print profits. It is a signal that the profit center has moved up the stack โ and that the stack above has a verification problem at its center.
The full PC analogy is more specific than the headline version. The PC price curve did not end in a golden age of cheap PCs. It ended in a trillion-dollar layer of software, services, and network effects that no one predicted from the hardware price alone. The AI analogue will end the same way. The applications that win will treat cheap intelligence as infrastructure, not as a product. And the crypto ecosystem that wins will be the one pairing cheap intelligence with a cryptographic integrity layer, because the cheaper the intelligence, the more expensive the lie, and the more valuable the proof.
The Goldman curve is a summary, not a schema.
It tells you the price fell. It does not tell you where the true cost falls, who absorbs the subsidy, whether demand elasticity holds, or what happens when false output gets cheap enough to flood every information channel. I read that curve the way I read a token's transfer function: find the point where the stated value diverges from structural reality, and position for that divergence. In 2017, the divergence paid me $12,000 in bounties. In 2022, it kept my capital out of collapsing lending protocols. In 2026, it points one direction for this industry.
The computation will keep getting cheaper. The verification will keep getting more necessary. And the people who confuse a falling price with a falling cost curve will keep subsidizing the people who understand the difference. Code is the only law that doesn't need a judge. But it still needs an audit โ and the audit has to start before the narrative does.