At 2:14 on a Tuesday morning in Melbourne, the most consequential document in technology is a pricing table. No whitepaper, no Discord server, no founder on a three-hour podcast. Just numbers. GPT-4o mini: $0.15 per million input tokens, $0.60 per million output. Three years earlier, GPT-3.5-Turbo charged $0.50 and $1.50 for the identical units. Same direction of travel. Different order of magnitude. Somewhere across the Pacific, Anthropic was running the same arithmetic in reverse โ Claude 3 Haiku repriced to $0.25 and $1.25, its mid-tier quietly re-platformed beneath it. Neither company issued a manifesto. They changed a number, and in changing it, they changed what it costs to pretend you are a company.
That is the hook. Now the part nobody puts on the chart.
I have spent enough hours tracing the ghost in the whitepaper's code to distrust any narrative that arrives pre-packaged with a moral. In late 2017 I spent three weeks inside the code of an ERC-20 token promising decentralized cloud storage. The economic model was logically broken โ I could demonstrate the collapse in a single afternoon with a spreadsheet. It raised $41 million anyway. What I learned then, and have never unlearned, is that technical correctness does not price a market; narrative cohesion does. The pricing table above is a narrative document disguised as an accounting one.
To read it properly you need the cycle behind it. Every crypto cycle has had a cost structure it was secretly fighting. 2017 fought the cost of trust โ intermediaries, banks, escrow, lawyers โ and proposed to replace them with a whitepaper's promise. 2020 fought the cost of access, and DeFi Summer turned yield farming into alchemy in the age of open protocols, where a liquidity pool was less a financial instrument than a group chat with a balance sheet. 2021 fought the cost of authorship, and NFTs went looking for the pixel that holds a soul. 2023 and 2024 fight the cost of cognition itself. Each cycle's slogan is different. Each cycle's actual question is identical: what is the cheapest way to produce the thing everyone now needs?
In August 2024 that question got answered out loud. Meta shipped Llama 3.1 405B โ 405 billion parameters, open weights, positioned openly as the strongest open model ever released and close to GPT-4o on the benchmarks people actually care about. Within weeks OpenAI cut the price of its small model by roughly 70 percent against its own 2023 baseline, and Anthropic pushed Haiku into territory it had never occupied. The crypto-facing press framed this as an aggressive push against open source. That framing is not wrong. It is just small.
Here is the part that matters for anyone holding a token whose value proposition is cheap inference.
The cost curve of intelligence has decoupled from the cost curve of the hardware that produces it. Over the same window in which API prices fell roughly 70 percent, the effective cost of an H100 โ dollars per unit of achievable throughput, adjusted for utilization โ fell somewhere in the range of 20 to 30 percent. Those two numbers cannot be reconciled by hardware alone. The gap is engineering. It is KV cache reuse across requests that never met each other. It is speculative decoding, which can lift throughput two to three times by letting a small draft model gamble and a large model confirm. It is dynamic batching that has learned to keep occupancy above the point where the math stops being embarrassing. It is Mixture-of-Experts sparsity, which means at inference time you are paying for a fraction of the parameters you advertised. The headline is a price war. The substance is an inference-efficiency breakthrough that the price war is merely making visible.
Generation-then-generation, this has been true for two years. GPT-4o arrived at less than half the unit price of GPT-4. Each Anthropic tier has reset its own price-performance ceiling. What has changed is the tier being contested: not the flagship, but the small model. That is a deliberate choice, and it tells you the strategy. The giants want the high-frequency, high-volume, low-complexity work to live on small models, and they want it to live there cheaply enough that nobody builds a company around replacing it. The flagship stays expensive. The floor drops out of the basement. A layered pricing pyramid is forming with the same logic as cloud egress: give away the doorway, charge for the hall.
Pull the crypto lens closer and the picture sharpens. On-chain inference was never competing with $0.15 per million tokens. It was competing with the sum of verification overhead, staking emissions, orchestration latency, and the token's own inflation. When I built the annotated sentiment dataset behind Human Pulse, the cost that mattered was never the raw forward pass. It was weaving trust into the immutable ledger โ proving who ran what, on which hardware, against which weights, under whose jurisdiction. Protocols that sell verifiability can survive a price war, because verifiability is not the line item OpenAI is discounting. Protocols that sell only compute are selling a commodity into a collapsing market.
Now do the arithmetic that actually decides this โ the total cost of ownership for self-hosting an open model.
Take Llama 3.1 70B. To serve it at production latency you need serious VRAM, or a multi-GPU rig with tensor parallelism and all the operational misery that implies. Add amortized GPU purchase or rental, power at Melbourne industrial rates or Ashburn data-center rates, cooling, networking, the engineer who understands why the CUDA out-of-memory error happens at 3am, the periodic re-quantization and re-validation, and the security surface of exposing inference to the public internet. Spread that across realistic utilization โ and realistic utilization for a small team is terrible, often single-digit percentages of capacity. Below roughly one hundred million tokens per month, the honest unit cost of self-hosted 70B inference exceeds what OpenAI now charges for a smaller model that outperforms it on most tasks. That sentence is the entire competitive event. Everything else is commentary.
The standard objection is autonomy. Data sovereignty. The right to run your own weights in a jurisdiction you choose. That objection is real and it is not going away, but notice what it has become: a compliance argument, not an economics one. Sovereignty was never cheap, and it was never supposed to be. The mistake was pretending it was also the cheapest option.
There is a second-order effect that the bear-market mind resists, and it is worth stating plainly because it is counter-intuitive and it is correct. Cheaper inference does not reduce total compute demand; it increases it. This is the Jevons pattern, and we already ran the experiment once. When block space got cheap enough to use casually, usage did not fall. Cheaper tokens mean more tokens: more agents running continuously rather than on demand, more retrieval loops, more speculative generation discarded before it reaches a user, more products that only pencil out if a phone call costs a fraction of a cent. The GPUs do not get idle. They get busier, the electricity bill grows, and the grid becomes the binding constraint somewhere around eighteen to twenty-four months out. Anyone who tells you a price cut is bearish for compute has not understood which curve they are reading.
And now the part my own readers will not enjoy.
The decentralized-compute narrative โ DePIN networks, marketplace aggregators, the long tail of idle-GPU protocols โ was priced on a single load-bearing assumption: that hyperscaler inference would stay expensive enough to leave a spread. Not merely a competitor to OpenAI; a cheaper alternative to it. That assumption is now under direct attack, and the attack is not rhetorical. When a centralized API undercuts your cost floor and beats your capability ceiling simultaneously, decentralized stops being a value proposition and becomes a subsidy. Some networks will survive on verifiable execution, on cryptographic proof of what actually ran, on workloads where jurisdiction matters more than latency. Most will discover that their token was doing the pricing all along, and that the token no longer agrees.
I wrote a ten-part series during the FTX collapse called The Silence Between Candles, mostly to keep myself from panic-selling my own convictions. The lesson from that winter transfers cleanly here: in a bear market, the only question that earns its keep is whether the thing you hold is bleeding or compounding. A protocol that loses forty percent of its liquidity providers in a quarter is not consolidating. It is dying with good posture. Apply that test to every AI-adjacent token in your watchlist and watch how quickly the decentralized-alternative language evaporates.
Here is the contrarian read, and it is the one I would stake reputation on. The open-source-versus-closed-source framing is a manufactured binary, and it is manufactured for the same reason liquidity fragmentation was manufactured in DeFi โ because a binary is easier to sell than a map. Real competition is not between philosophies. It is between capital blocs with cloud infrastructure attached. Microsoft funds OpenAI and rents it Azure compute. Amazon funds Anthropic and rents it Trainium silicon, which means part of Anthropic's price cut is being paid for by a company that also sells the alternative to the thing Anthropic is cutting. Google sits outside the fight entirely, having already priced Gemini Flash below the floor, subsidized by TPUs it builds itself and distribution through Workspace and Android that no API price can replicate. Chasing the myth through the ledger's fog, you find that open source is not the opponent. Open source is the veil the blocs hold up while they maneuver against each other. Meta does not need Llama to win. Meta needs Llama to make sure nobody else wins outright.
There is an older parallel worth sitting with. When the spot Bitcoin ETF was approved, the ticker survived and the cypherpunk premise did not. Peer-to-peer electronic cash became a settlement layer for institutional balance sheets, and the price rose anyway. Decentralized AI is walking the same corridor. The tokens will keep trading. The governance will keep happening in Telegram polls with eleven participants.
What I am watching now is not the next price cut. It is the moment โ somewhere in the next twelve to eighteen months โ when an open model closes to within ten percent of the frontier at half the inference cost, and the pricing table that started this essay stops being a strategy document and becomes a confession. When that number arrives, which side of the ledger will your conviction be standing on?