Capability brief · Independent · Evidence-first

NVIDIA lets machines read its story. It hasn't yet let them verify the hardest chapter.

NVIDIA is one of the few large companies whose public surface an agent can actually navigate. Findings below, hardest evidence first: every one re-checkable in a single command. The strong baseline is credited in detail, because it earns it; the gap is precise and sits exactly where the world is looking.

Method. Passive HTTP only: curl header/file GETs of standard public paths, a soft-404 control, and WebSearch/WebFetch of NVIDIA's own report + primary sources for figures (labeled reported). No authentication, no scanning, no enumeration beyond standard well-knowns. Captured 2026-07-14. Not affiliated with or endorsed by NVIDIA.

Baseline: credited first, in detail

NVIDIA is genuinely ahead on agent legibility. www.nvidia.com/llms.txt (200, text/markdown) is a real index that fans out to a maintained hierarchy: docs.nvidia.com/llms.txt (200, 7.7 KB, self-dated "Last updated 10 July 2026"), developer.nvidia.com/llms.txt (200, ~28 KB), plus locale and blog files. developer.nvidia.com/.well-known/agent-card.json is a real A2A (Agent2Agent: the protocol AI agents use to discover and call each other) agent card (200). And robots.txt enumerates 50+ named AI crawlers and explicitly Allows them the markdown lane. This is not a "you're behind" brief, it's "you built the reading room; the verification vault is missing."

The bottom line, in plain English

No jargon. NVIDIA already built the best agent-reading room we've audited, every fix below just extends that lead to the one story the whole world is now scrutinizing. Here is what each one is worth.

A verified retirement receipt per unit of compute → the most-watched question in tech gets an answer better than “trust our PDF”

Your Scope-3 footprint doubled to 6.9M tonnes and lives only in a self-authored PDF; an on-chain retirement receipt per unit of compute turns that story into something an auditor, regulator, or agent can independently confirm.

A door on the corporate host → the agents and regulators asking about your climate story actually reach you

developer.nvidia.com is agent-ready, but the marketing host where the sustainability story lives returns 403 to every agent probe; opening it means the reading room you already built finally has a door on the building where the hard questions land.

Fix the markdown skill your agent card promises → agents that trust NVIDIA don't hit dead ends

Your A2A card advertises a universal /{path}.md skill that only partly resolves, so an agent following it hits 404s on real topics; completing it protects the very legibility NVIDIA invested in.

A CSP cleanup → no test-environment hostnames left exposed in production

Your production frame-ancestors list still carries UAT/QA hosts and one malformed token; a config sweep removes an unforced signal without touching anything a user ever sees.

Finding 1: the lead (credit, with the receipt)

Provable · NVIDIA's agent posture is real

A maintained llms.txt hierarchy, an A2A agent card, and an explicit AI-crawler markdown lane.

The evidence for the credit, so it's not just praise:

$ curl -s -o /dev/null -w "%{http_code} %{content_type} %{size_download}B\n" https://docs.nvidia.com/llms.txt 200 text/plain 7703B # header: "Last updated 10 July 2026" $ curl -s https://developer.nvidia.com/.well-known/agent-card.json {"name":"NVIDIA Developer Site Agent", "skills":[{"id":"site-context",...}, {"id":"page-markdown","endpoint":"https://developer.nvidia.com/{path}.md"}]} $ curl -s https://www.nvidia.com/robots.txt | grep -A2 'anthropic-ai' User-agent: anthropic-ai ... (50+ AI agents listed) Allow: /*.llms.txt$ Allow: /*.md$
Why it matters: most companies force agents to scrape rendered HTML. NVIDIA hands them curated markdown and an identity card. That instinct is exactly the one this concept builds on, it just needs to reach the sustainability story.
Re-check: curl -s https://www.nvidia.com/llms.txt · curl -s https://developer.nvidia.com/.well-known/agent-card.json

Finding 2: the corporate host is dark to agents

Provable · where the footprint lives has no agent identity

Every agent well-known on www.nvidia.com returns 403; ai.txt and sitemap.xml soft-404.

The developer host is agent-ready. The corporate/marketing host (where the sustainability and investor story lives) is not: the entire /.well-known/ tree is blocked at the edge, and ai.txt/sitemap.xml return the SPA 404 shell (byte-identical to a random control, so they're soft-404s, not real files).

$ for p in agent-card.json ai-plugin.json security.txt; do curl -s -o /dev/null \ -w "$p %{http_code}\n" https://www.nvidia.com/.well-known/$p; done agent-card.json 403 ai-plugin.json 403 security.txt 403 $ curl -s -o /dev/null -w "ai.txt %{http_code} %{size_download}B\n" https://www.nvidia.com/ai.txt ai.txt 404 298775B $ curl -s -o /dev/null -w "control %{http_code} %{size_download}B\n" https://www.nvidia.com/this-is-a-control-xyz123.txt control 404 298775B # identical size → ai.txt is a soft-404 SPA shell
Why it matters: when an agent (or a regulator's tool) asks "who speaks for NVIDIA on climate, and can I verify a claim?", it lands on a 403 or an HTML shell. The reading room NVIDIA built doesn't yet have a door on the building where the hardest questions are asked.
Re-check: curl -sI https://www.nvidia.com/.well-known/agent-card.json

Finding 3: the agent surface is uneven

Provable · partial coverage + a caching wrinkle

The agent card advertises a universal markdown skill that only partly resolves.

The developer agent card advertises a page-markdown skill at https://developer.nvidia.com/{path}.md. Coverage is partial: one representative path works, two return 404 (HTML shell). Separately, developer.nvidia.com/llms.txt returned 0 bytes on a first cold GET then the full ~28 KB on immediate retry: an edge-cache cold-miss, reported honestly as a reliability note, not a broken file.

$ for p in cuda-downloads blog deep-learning-ai; do curl -s -o /dev/null \ -w "$p.md %{http_code} %{content_type}\n" https://developer.nvidia.com/$p.md; done cuda-downloads.md 200 text/markdown blog.md 404 text/html deep-learning-ai.md 404 text/html $ curl (1st) developer.nvidia.com/llms.txt → 200, 0B ; curl (retry) → 200, 28803B
Why it matters: an agent that trusts the agent card will call /{path}.md and hit 404s on real topics, eroding the very legibility NVIDIA invested in. Small, fixable, and worth naming precisely rather than glossing.
Re-check: curl -s -o /dev/null -w '%{http_code}\n' https://developer.nvidia.com/blog.md

Finding 4: the headline gap

Provable (reported figures) · the footprint is PDF-bound + self-attested

The #1 question over the AI buildout is answered in a PDF, by NVIDIA, about NVIDIA.

NVIDIA's FY25 Sustainability Report is a downloadable PDF. It reports real progress: 100% renewable electricity for owned/operated offices and data centers (genuine, credited), but also that Scope 3 emissions jumped from 3.5M to 6.9M MT CO₂e (purchased goods & services +~2.8M). None of it is machine-verifiable, and there is no per-unit-of-compute retirement receipt an outside party can check. In an AI buildout the whole world is scrutinizing, the proof is a self-authored document.

# primary source (fetchable PDF): https://images.nvidia.com/aem-dam/Solutions/documents/NVIDIA-Sustainability-Report-Fiscal-Year-2025.pdf # reported: Scope 3 3.5M → 6.9M MT CO2e ; 100% renewable (owned ops) ; FY26 goal: engage # suppliers = 67% of Scope-3 cat-1 to adopt science-based targets
Why it matters: "trust our PDF" is the weakest possible answer to the most-watched question in tech. A verified, on-chain ecological retirement receipt per unit of compute, the EcoWealth/Regen pattern, is the object that turns the Scope-3 narrative into something an auditor, regulator, or agent can independently confirm.
Re-check: open the PDF URL above · WebSearch "NVIDIA FY25 sustainability report Scope 3 6.9 million"

Finding 5: hygiene (minor)

Provable · a small production-CSP note

The marketing CSP frame-ancestors list carries lower-environment hosts and one malformed entry.

www.nvidia.com's content-security-policy: frame-ancestors allowlists UAT/QA CMS hosts (cms.nvidia-docs-uat…brightspot.cloud, …qa3…, docs.nvidia-docs-uat…) alongside production origins, and ends on a truncated/malformed token (https://nvfile;). Low stakes: the kind of thing a CSP cleanup sweeps up.

$ curl -sI https://www.nvidia.com/en-us/ | grep -io 'frame-ancestors[^;]*nvfile[^;]*' frame-ancestors ... nvidia-docs-uat.lower.k3.m1.brightspot.cloud ... https://nvfile
Re-check: curl -sI https://www.nvidia.com/en-us/ | grep -i content-security-policy

The ask

One verified retirement receipt. A few dollars. That's the whole first step.

Retire the compute footprint of a single AI session to a verified on-chain Regen Network certificate (pocket change at NVIDIA scale) and keep the certificate. It gives the Scope-3 story its missing verifiable object: a unit of compute, its ecological cost, addressed and independently checkable.

Or have any AI agent make one keyless call: x402 (a web-native way for AI agents to pay per use) POST vealth.net/.well-known/vealth/work-tco2e ($0.005) or /work-packet ($0.10), and fund one place-based restore packet near an NVIDIA data-center community. Deliverable: the on-chain proof receipt.

Why this and not a meeting: it ties directly to NVIDIA's own stated commitments (100% renewable, supplier science-based targets) and hands the next sustainability report a receipt instead of a claim. Proof first, ask second.
Provisioned by EcoWealth · Digital Provisioning Protocol · provenance‑sealed