The most awkward number in AI this week is a flat line: usage keeps climbing, but Ramp says spending has stopped following it. That lands just as Anthropic and OpenAI are trying to sell the world on trillion-dollar valuations.

1. AI usage is soaring. The revenue isn’t following.

Ramp says token spend grew just 1.5% across August and September, after jumping 42% in July. Token volumes kept rising anyway. That is a lot more AI for roughly the same money, which is wonderful for buyers and an awkward fact for companies seeking trillion-dollar valuations.

Ramp economist Ara Kharazian blames price competition and customer discipline. In September OpenAI halved API prices for GPT-6 Sol and Luna, and Anthropic cut prices on Fable 5.1 and Opus 5.5. Anthropic said in a filing that its cuts shaved about $600 million off revenue growth. Customers are also getting thrifty, with budgets, defaults that block expensive models, and routers that push work to cheaper ones. Notion’s Akshay Kothari put it as “tokens go brr. Costs go ahhh.“

The deeper problem is concentration. Ramp says almost all spend comes from venture-backed tech firms, with the top 1% near $10,000 per employee against a median of $12 to $15 a month. A car dealership is unlikely to join that top 1%. October did rebound 8.9%, so this isn’t a verdict. But per TechCrunch, Bloomberg reports Anthropic’s potential valuation fell from $2 trillion to $1.2 to $1.3 trillion, and a Register headline says its revised IPO filing warns 12x revenue growth is unlikely to continue.

TechCrunch · The Register

2. Jensen Huang gets a congressional summons, and the details are thin

Ars Technica reports that lawmakers want answers from Jensen Huang about his ties to an alleged Chinese intelligence asset. That is the supported record so far: a summons and an allegation, not a finding of wrongdoing.

It still matters because of whose name is on it. Nvidia sits underneath most of the AI buildout, and any question about its chief executive and China draws attention to export rules and customers. But I’d resist drawing that line yet. Who the person is, what the ties were and what Congress wants from Huang aren’t in the material available. The next real signal is what lawmakers say after he responds, or whether he does.

Ars Technica

3. Flock’s problem isn’t the cameras. Nobody voted for them.

At least ten local governments and police departments have moved since summer to restrict or drop Flock, according to TechCrunch. Monroe, Newburgh, Richmond, New Bedford and Lansing are among them. League City, Texas, voted this week to remove its license plate readers. That’s a notable setback for a company with more than 100,000 cameras in use.

The revealing detail is how Flock got in. The EFF’s Adam Schwartz says police departments often adopt it with little council oversight, and South Windsor, Connecticut, reportedly went live without public warning or a vote, then voted to remove the cameras in October. Councils control procurement, and activists are now using that.

Flock points to solved crimes, says it doesn’t sell data or contract with ICE, and has cut default retention from 30 days to seven and required case codes for searches. The cases of officers using it to stalk ex-partners and reported federal access are harder to answer. Cleveland shows the likelier outcome for many buyers: it renewed, restricted immigration searches and cut off a fusion center, while reviewing rival bids. Trust is now part of the product.

TechCrunch · The Register

4. The UK wants a way in when AI agents slip the leash

The Register reports that UK ministers want powers to investigate when AI agents are “let off the leash.” I can’t confirm from the material here what those powers include, so treat claims about seizures or inspections as unverified. The framing is still telling: regulators are preparing for the case where an operator has lost track of what its agent is doing.

Cloudflare’s essay explains why that case is plausible. A hosted agent is an intermediary, and the website sees the agent rather than the person behind it. Sites can’t tell users apart, ban one abuser without blocking everyone, or know who answers for a bad purchase. Cloudflare argues the web needs identity, reputation and accountability, and backs signed agent requests (Web Bot Auth) and privacy-preserving rate limiting (MOLE).

The practical lesson for agent builders is that visibility is becoming a product requirement. If you can’t say who your agent acted for and what it did, a regulator and a site operator will both ask.

The Register · Cloudflare Blog

5. GitHub rebuilt Git storage for 5,000 agents pushing at once

GitHub says the busiest repositories today see at most dozens of writes a second. Its new architecture, built as “Project Rubicon,” handled 7,200 pushes a second from 5,000 concurrent agents in internal testing, with reads staying low-latency. That is a deliberate overbuild for a workload GitHub expects within a year.

The key insight is that very little of a push needs agreement. Git objects are immutable, so only reference updates must be ordered. GitHub put packs in Azure Blob Storage and references in a small replicated log called Ref-CAS, and moved compaction onto separate nodes. Its earlier blob-only designs topped out in the mid-hundreds of writes a second. Git itself stays Git for users.

Note what this doesn’t show. It is GitHub’s own test, and it addresses storage, not whether CI queues or human reviewers can absorb that volume. The security market is already betting on agent access, with Palo Alto reportedly in talks to buy Clutch for $500 million, about 50 times its roughly $10 million ARR, per a TechCrunch source.

GitHub Blog · TechCrunch