The question this column keeps circling — what does a completed task cost, and who gets to meter it — produced two answers this week that point in different directions. NVIDIA published research arguing the harness, not the model, is where frontier agent performance comes from, reporting its agent system carries Claude Opus 5 from roughly 30% to a perfect score on the ARC-AGI-3 public set. Real corporate billing data from more than 70,000 businesses says the market answers with its wallet: adoption tracks price, and the cheapest thing that completes the task owns the company card. Meanwhile the open layer’s neutral plumbing kept being absorbed — the router, then the registry, now the database your agent queries, each one gaining a captable. We fired a breaking note this morning on Anthropic anchoring its post-IPO compute to a 2024-founded neocloud; here is what else the day held.
NVIDIA wants to own the harness, because the harness is now the model
What happened. NVIDIA’s research team reports that its Agentic Variation Operators architecture (AVO, arXiv 2603.24517 — an agent system built around persistent memory and a supervisor that interrupts stagnation) scores 100.00 RHAE on the ARC-AGI-3 public set, solving all 183 levels across the 25 environments in 6,624 actions versus the 7,542 VISTA reports, roughly 12% fewer. The post is careful about scope: the ~30% “model baseline” is ARC Prize’s own score of Claude Opus 5 at high reasoning effort under a different setup, and NVIDIA says explicitly this is not a controlled ablation and covers only the public set, not the private ones. The same architecture ran unattended for seven days on GPU-kernel optimization, committing 40 kernel versions and posting up to 10.5% over FlashAttention-4 on DGX B200.
Why it matters. Read past the number to what the vendor is actually claiming: “evaluating a model is not the same as evaluating an agent,” and the surrounding system — memory, supervision, tool loop — is where long-horizon capability comes from. Whether or not 100.00 holds up under outside measurement, that claim is the strategic tell. If model identity matters less than system design, the layer an operator should shop for is the harness, and the vendor that gets to claim the harness also gets the telemetry, the memory, and eventually the bill. Treat the headline like the Vera Rubin 30x figure: it is a vendor-run research result on a public eval set, with fine-print caveats NVIDIA itself prints, nobody outside NVIDIA has re-run it, and a preprint is not a product under load. The genuinely useful part is the mechanism list — persistent memory and a supervisor that redirects a stalled agent — because those are exactly the two components that die first in every long-running deployment I have seen.
Source: developer.nvidia.com
Price is a feature: corporate billing data says the cheapest completed task wins
What happened. The Financial Times reports that Anthropic’s best model is struggling to attract users as cheaper tools thrive, per people with knowledge of the matter, and the reporting leans on Ramp’s AI index — an adoption measure built from real billing transactions at more than 70,000 businesses rather than surveys. Ramp’s index puts paid AI adoption at 55.7% of U.S. businesses in July, against the Census survey estimate of 21.6%, and Simon Willison’s read of Ramp’s Anthropic spend data supports the idea that Fable’s premium price has left it the least-adopted model in the line (Opus 5 shipped July 24).
Why it matters. The operating economics divide becomes a meter you can read. Adoption is not tracking benchmark position; it is tracking the cheapest thing that completes the task, and the gap between a billing-based index (55.7%) and a survey (21.6%) is the same gap this column predicted when open weights started setting prices: real procurement overcounts the cheap, good-enough stuff and undercounts what people claim in surveys. For an operator the lesson is concrete — the routing layer you build should reprice to the cheapest completion, which is why the router sale to Stripe matters, and it is why a “best” model is a feasibility claim, not a deployment default. Ramp is a credit-card view and will miss datacenter-scale multi-tenant contracts, but as a proxy for who actually gets paid, it beats every leaderboard.
Source: ft.com, ramp.com, simonwillison.net
The open layer’s data plane just got a landlord, like the router and the registry
What happened. AWS signed a definitive agreement to acquire DuckLabs, the Amsterdam company behind DuckDB, the embeddable open-source analytics database that apps increasingly ship as their query engine. AWS says it is not acquiring the DuckDB open-source project — it stays MIT-licensed under the independent DuckDB Foundation — but founders Hannes Mühleisen and Mark Raasveldt join AWS and keep leading the project’s technical direction. Financial terms were not disclosed; the deal is expected to close shortly.
Why it matters. The Hugging Face arc has a data-plane sequel. The model router went under a balance sheet, the model registry is being shopped, and now the embedded database your agents query reports to the same company whose object store you already host on. “Open, MIT, free” today is fine — AWS is buying the team, the roadmap, and a default on-ramp into S3, Athena and SageMaker, which is the whole strategic value. The failure mode for an operator is not a license change; it is slow drift where the neutral engine gets optimized for its parent cloud, and an eval that says “just pull the OSS, it’s neutral” quietly needs a captable column. DuckLabs and AWS have collaborated since early 2025 (S3 Tables, SageMaker Lakehouse); being absorbed is the endpoint of that happy path.
Source: aboutamazon.com, aws.amazon.com
The Rest
- Token-in, token-out done right — Hugging Face’s TITO guide pins down a fixable bug class in agentic RL: never re-encode tokens you’ve decoded, and keep the chat template prefix-preserving for tool messages. The re-tokenization drift it kills costs GPU time and training quality on every long agentic trace. huggingface.co
- IBM’s ALTK-Evolve: agents learn from their own trajectories — it distills behavioral guidelines from past runs and reinjects them at inference, no weight updates and no labels. The finding worth keeping: the right memory dose is tier-dependent — strong models want the full guideline set, weaker models a compact core, saturated models show no gain. Memory sizing is becoming a per-workload engineering decision. huggingface.co
- IAI and Thales to build Nigeria’s next comm satellites, ~$2B reported — NIGCOMSAT expects to buy two spacecraft under Federal Executive Council approval: IAI’s NigComSat-2A (launch 2028) and Thales’ 2B (2029). The value is undisclosed and estimated at around $2 billion, and IAI has not commented — flagged as reported, not confirmed — but it is a meaningful win for Israel’s space division after its Amos-6-era drought. calcalistech.com
- United resumes San Francisco–Tel Aviv in March 2027 — 3×/week from March 28 with more business-class seats than any carrier on the route, while El Al gets there first with an SFO launch on October 25, 2026. The Bay Area–Israel corridor reconnecting is a quiet signal for the diaspora-tech bridge. calcalistech.com
What I’m watching
A scorecard is queued on this morning’s Nscale breaking note: does the $45 billion, ~460MW, late-2027 ramp survive contact with a 2024-founded operator’s financing, and does Anthropic confirm the term on the roadshow. On the harness claim, I’m watching whether anyone outside NVIDIA runs ARC-AGI-3 with their own system — external arbitration is the only thing that makes a perfect score mean something for the rest of us. And on the open layer, DuckDB joined OpenRouter and (reportedly) Hugging Face behind corporate owners; the open question is whether “open but neutral” survives a captable, one layer at a time.