AI x Crypto

AI Crypto Sector Up 54% as Agents Get Graded on Reality

Grayscale says AI crypto jumped 54% in September, while Microsoft's ThinkingBox grades agents on database state — not vibes. Plus Bitget's $170M mystery.

  • AI x Crypto
  • Markets
  • Agents
AI Crypto Sector Up 54% as Agents Get Graded on Reality

The AI x Crypto Trade Is Back On, Allegedly

Grayscale says the AI crypto sector jumped 54% in September, with NEAR up 183% leading the pack, per finance.biggo.com. That's a number that will get recycled in a lot of pitch decks this week.

Meanwhile, Arthur Hayes is already calling the end of the AI bubble, predicting a trillion-dollar rescue (Bitcoin News). So: sector up 54%, bubble ending, rescue incoming. Pick your narrative, the chart doesn't care which one you picked.

A 54% monthly move in a sector whose tokens mostly do not yet have revenue is not a signal. It's a mood ring.


Agents Get Graded on What They Actually Do

The most substantive AI story of the cycle is not a token. Microsoft and Hugging Face shipped ThinkingBox, a benchmark that grades AI agents on the terminal backend state and side effects they leave behind — not the sentences they generate.

  • Across 507 stateful business workflows, each run 20 times against various LLMs.
  • The example: an agent handles a stuck appliance delivery with nine well-formed tool calls, then closes the ticket as "resolved" — while the carrier exception is still open and the customer never got an answer.
  • A grader checking tool calls sees nine good ones. The database disagrees.

That gap is the whole point. It's also, quietly, the exact gap that decentralized agent networks will have to solve before anyone trusts an autonomous wallet.

Related research lands on the same theme: an arXiv paper on "early posterior collapse" finds LLMs commit to an early interpretation and treat later clarification as extra context rather than a correction. Another finds chain-of-thought is only "load-bearing" on hard tasks — on easy ones, models silently bypass their own reasoning. Where the trace is easy to read, it carries little signal. Where it matters, errors propagate before a monitor can intervene.

Everyone building "verifiable AI agents" should read the ThinkingBox post twice. Verification of words is theater. Verification of state is the product.


Inference Gets Cheaper and More Crowded

Hugging Face keeps stacking serverless inference providers:

  • Fireworks.ai joined the Hub, serving DeepSeek-R1 and others directly from model pages.
  • Hyperbolic, Nebius AI Studio, and Novita were added, joining Together AI, SambaNova, Replicate, and fal.
  • A separate HF post breaks down the economics of running 1B+ classifications or embeddings per day — roughly English Wikipedia 144x over — and how to not set your cloud bill on fire.

The direction is clear: inference is commoditizing into a routing layer. The margin moves to whoever owns the cheapest GPUs, not the best model page.


Meanwhile, in Crypto Plumbing

  • Bitget: more than $170 million moved from wallets labeled as Bitget's to a single fresh address, spanning ETH, USDT, USDC, AVAX, and BNB. The CEO acknowledged the incident on X; it's still unclear whether it was a breach or an internal operation (The Block).
  • Tether says less than 0.034% of group assets sit at EQIBank, the Dominica-licensed bank claiming US seizures took ~80% of its financial assets — under ~$63.8M against $187.75B in Q2 reserves (Unchained, CoinTelegraph).
  • Polymarket got sued by New York over alleged unlicensed gambling, and sued back in federal court the same day (Unchained).
  • Bitcoin ETFs took in $191M Thursday, pushing the six-day streak past $2.8B, though daily inflows are down 81% from Monday's $999M peak (CoinTelegraph).

The Rest of the Noise

  • CoinMarketCap launched an AI Agent Hub — a directory play, not a product play.
  • Microsoft bundled chat, coding, and agents into a Copilot "super app" and renamed Scout to Autopilot (The Verge via Techmeme).
  • Databricks acquired spreadsheet startup Row Zero to feed its AI coworker Genie (Axios via Techmeme).
  • The NSA told lawmakers it's spending billions this year testing AI models, while proposals for a US AI regulator were estimated at just $20M–$40M/year (Washington Sun via Techmeme).

Closing Take

The AI crypto sector's 54% September is the easy headline. The harder one is that the industry's most credible new benchmark says your agent's confident summary is worth nothing if the database disagrees — and almost nobody trading AI tokens this week has read it.

Not financial advice. Side-Eye Signals aggregates and snarks; verify everything yourself.