Nemocore.ai fuses NVIDIA Nemotron with Fastly's global network to treat inference the way the edge already treats content: classify it, diff it, route it, cache it, purge it, and protect it. Same primitives, new payload.
Agent workloads resend near-identical context turn after turn — the same repo, the same policy doc, the same customer record — paying full token price every time, with no consistent policy for which model or region should even see it.
GET /service/{id}/diff/from/{v1}/to/{v2} — line-by-line config diffs on demandNemocore sits in front of any model router — Nemotron, a self-hosted model, or a third-party vendor — and applies edge discipline to the data before a single token is spent. Click a layer to see it up close.
Every chunk of context is tagged — public, internal, confidential, restricted — by pattern rules plus a small local classifier that never leaves the edge node. This is the WAF, pointed at meaning instead of packets.
runs at the edgeContent-addressed, hash-based store — the same lineage as Fastly's Diff API, applied to prompt context instead of service configs. If a near-identical version of this context was already sent and approved, only the delta goes out.
content-addressedRoutes by sensitivity tier first, cost and latency second. Restricted data never sees a vendor option. Confidential data only reaches vendors under a signed DPA with zero data retention.
tier-awareContext expires on a schedule keyed to its classification. Confidential and restricted data can be purged immediately post-task — instantly, self-serve, the same way Fastly purge replaced the old ticket queue.
policy-drivenEvery classify → route → purge event, timestamped. The answer to "did anything confidential leave the building, and where did it go" — on demand, not by reconstruction.
compliance-readyNemocore inherits Fastly's Next-Gen WAF and Bot Management outright — every request into the gateway gets app/API-grade protection before it ever reaches a model.
Evaluates request context and execution logic instead of static regex rules, so the gateway blocks prompt-injection and payload-smuggling attempts with near-zero manual tuning — and gets smarter through Fastly's Network Learning Exchange, threat intelligence shared anonymously across tens of thousands of agents.
Automated scraping and credential-stuffing patterns aimed at a model endpoint look a lot like the bot traffic Fastly already fingerprints at the edge. Nemocore applies the same detection in front of the router — before a bot ever gets to consume a single token.
Once sensitivity is settled, Nemocore's router optimizes across Nemotron variants and third-party models by cost, latency, and region — the same origin-selection logic Fastly already runs for HTTP traffic, applied to inference.
Local-only Nemotron instance. Never leaves the origin network. No vendor option offered.
Vendor with signed DPA + zero data retention only. Nearest compliant region wins on latency.
Full cost/latency optimization across Nemotron and contracted vendors, any eligible geography.
Nemocore is designed to be adopted incrementally, starting inside infrastructure you already run.
Every infrastructure company has an origin story. Ours involves a lot of rejected domain names.
Our story began like so many before us.
Sure, we could have called it Claudeflare. But something about a certain someone's relentless obsession with mentioning their founders are lawyers gave us the willies. And we never really understood what a "flare" was, beyond an SOS safety tool and/or an unwanted occurrence on the body — and we figure the AI industry's logo obsession already has that whole area of content covered.
We even bought the Claudeflair domains and commissioned a genuinely snazzy logo. Then our own relentless business acumen and hotshot team of lawyers told us to get back to work.
Try getting "Traingle" past autocorrect. We dare you. None of them felt quite as natural as Nemocore.
And while Nemotron and NVIDIA are central technologies and co-design partners, we are by no means bound to either. We actively encourage customers to BYO models and edge providers to the SlopChopShop for even greater savings and modularity.