Cloudflare (Workers AI)

Thesis

Cloudflare is converting Workers AI from a hosted-model catalog into a first-party AI platform and agentic-computing substrate. Its first self-trained models — Clef and Clef-flash, open-weight decision models released under Apache 2.0 — plus a new RL fine-tuning service mark the shift from inference hosting toward model training W1W2W4. Hiring clusters on the AI/agent stack (Workers AI, AI Gateway, ML, Agentic Commerce) and on added ML-infrastructure talent from Ensemble AI E11E14E21E57E58W5W6, while public framing centers the "agentic Internet" and AI-agent governance P5P7. This AI push runs on top of Cloudflare's durable network/security franchise: Tbps-scale DDoS mitigation, Gartner SASE/SSE Visionary recognition, and post-quantum cryptography P4P5E1.

Signal desks

Hiring

• AI/ML buildout: Senior Systems Engineer, Workers AI E11; Senior ML Engineer E21; Senior Systems Engineer, AI Gateway E14; Workers AI Engineer (Austin, $145k–$195k, hybrid) W6. Cloudflare is "deepening our investment in AI with the addition of team members from Ensemble AI, focusing on machine learning infrastructure and efficiency" W5.
• Agentic commerce: Software Engineer and Senior Software Engineer, Agentic Commerce E57E58.
• Workers/runtime & storage platform: Workers Runtime E15; Workers Deploy & Config (Principal + SE) E9E10; Durable Objects E17; Queues E19; Hyperdrive E20; R2 Gateway E5E6; R2 Metadata E16; Basin Pipelines E12; Basin Platform E13; Storage Platform E22; Distributed Databases E18; Cloudflare Tunnel E49.
• Security/threat intel: Software Engineer (Threat Intelligence), Cloudforce One E54.
• GTM/international: Central & Eastern Europe customer engineering/sales (Russian-, Hebrew-speaking, Digital Natives) P11P15E37E38E39E40; Senior Named Account Executive, Saudi Arabia (Dubai) P12E47; Principal, GTM Automation RevOps (building "AI agents and automation" for the revenue org) P17E33; Program Lead Strategic CDN Accounts GTM E36; Senior Events Manager P14E24.
• Infrastructure/ops: Data Center Network Operations Engineer (US hubs; Lisbon/London) P13P16E34E35; Logistics Operations Specialist E43.
• People ops: Workday Reporting Specialist / Analyst (data governance, workforce analytics) P3E27.

Location read: Austin emerges as the Workers AI engineering hub W6P14; San Francisco for RevOps P17; London/Amsterdam/Lisbon for CEE customer engineering and network ops P11P15P16; Dubai for Middle East sales P12E47; multiple US DC/network hubs P13.

Forks No cited evidence in this pack. (The "forge" items are auto-generated OpenAPI artifacts from cloudflare/api-schemas, not forks of upstream projects P8P25E4.)

Releases

• Models: Clef and Clef-flash open-weight decision models on Workers AI + Hugging Face, Apache 2.0, plus an RL fine-tuning service W1W2W4.
• Agent/compute runtimes: @cloudflare/computer@0.7.0 (experimental onInterrupt restart/replay, ctx.interrupt) P20E48; agents@0.30.1 (PiHarness onWake recovery fix) P23E42.
• Workers runtime: workerd v1.20261025.1 (continuation capture weak handles, WebAssembly shims, Rust JSG resources) P9E2; workerd-tools wpt tag P10E3.
• Developer tooling: @cloudflare/vite-plugin@1.65.0 (removes experimental.newConfig; Bun binding error handling) P19E52; workers-py v1.17.9 E53; cf@1.0.0-beta.21 (stderr/stdout split, telemetry) P21E46; speedtest v1.15.0 P18E50; lol-html v3.0.2 P22E44; networkquality-rs v0.4.0 P24E41; forge openapi auto-releases P8P25P26P27P28E4E25.

Talking

• Clef launch narrative ("first models trained by the Cloudflare Workers AI team") W1; third-party coverage framing Clef as "probabilities for agents instead of text generation" W2; a spec table with Qwen3.8-27B / Qwen3.5-9B backbones W4.
• Post-quantum cryptography: GeoKey Manager for geographically fenced keys against harvest-now/decrypt-later attacks E1.
• Agentic Internet economics: Content Independence Day one-year report (2.5B genAI users, collapse of open-web search) P7.
• AI-agent security framing: SASE/SSE Visionary post — "first SASE platform to rein in MCP server sprawl, natively governing AI agents" P5.
• Threat landscape: DDoS Threat Report H1 2026 (935 attacks >1 Tbps; DNS floods) P4; Project Galileo civil-society attack report P1.
• Infrastructure/engineering: Origin Range Requests (Pingora cache rebuild, R2 p99 latency cut in half) E29; South Korea ISMS compliance at scale E60.

Shipping

The clearest shipped artifacts are Clef and Clef-flash: open-weight decision models (Apache 2.0) hosted on Workers AI and Hugging Face, shipped alongside a new RL fine-tuning service W1W2W4. Clef is a 27B (Qwen3.8-27B) precision model at $0.24/M input tokens; Clef-flash is a 9B (Qwen3.5-9B) latency model at $0.09/M W4; platform pricing is $0.011 per 1,000 Neurons with a 10,000/day free allocation W3. In parallel, Cloudflare shipped agent/compute and runtime updates: @cloudflare/computer@0.7.0 adds replayable runs and interrupts P20, workerd continues near-daily releases P9E2, and the workers-sdk Vite plugin hit v1 with Wrangler-only config P19.

Research themes

Cloudflare's public research narrative centers on a decision-model paradigm — models that return probabilities over typed answers instead of generating text — positioned for agent action (route, block, escalate) with no free-form parsing and no reasoning tokens W1W2W4. Adjacent themes: RL fine-tuning as a service W1; post-quantum key management E1; and replayable/interruptible agent execution P20. Note on evidence quality: two items in this pack carry Cloudflare-titled metadata but non-Cloudflare body content — P2 is an OpenAI–Hugging Face incident technical report and P6 is the Stanford AI Index 2026 — so they are excluded from Cloudflare-specific research claims P2P6.

Hiring & scaling

The hiring data shows a dual-track buildout. First, an AI platform track: Workers AI systems engineers, ML engineers, AI Gateway, and the Ensemble AI talent addition indicate investment in training/inference infrastructure rather than only inference hosting E11E21E14W5W6. Second, an agentic-GTM track: Agentic Commerce engineering roles E57E58 and a Principal GTM Automation RevOps role that "build[s] the AI agents and automation" for the revenue organization P17 signal commercialization of the agentic Internet thesis P7. International expansion continues via CEE and Saudi Arabia/Dubai customer-engineering and sales roles P11P15E39E40P12E47, while data-center network ops roles point to ongoing global network maintenance P13P16.

Category implications

• Product: Workers AI is differentiating from pure inference aggregation by shipping first-party open-weight models and an RL fine-tuning service — turning the platform into a model-training destination, not just a model host W1W4.
• Infrastructure: Training Clef/Clef-flash and offering RL fine-tuning implies GPU/compute capacity beyond serving, reinforced by Workers AI and ML engineering hires W1W6E11E21.
• Strategy: Cloudflare frames the "agentic Internet" as a market shift and positions its connectivity cloud to govern AI agents and MCP server sprawl, extending its security franchise to the agent era P5P7.
• Research: The decision-model approach targets structured, latency-sensitive agent decisions, a deliberate alternative to generative LLMs W1W2.
• Hiring: The AI/agent hiring wave (Workers AI, AI Gateway, ML, Agentic Commerce, Ensemble AI) is the leading indicator of this pivot E11E14E21E57E58W5W6.
• GTM: RevOps automation and agentic-commerce roles indicate AI is being applied to Cloudflare's own go-to-market and to enabling AI-driven commerce on its platform P17E57E58.

Traction highlights

• Clef launch traction: 457 Hugging Face likes (Clef) and 161 (Clef-flash) observed at launch W4.
• Model economics published: Clef $0.24/M input tokens, Clef-flash $0.09/M W4; Workers AI $0.011 per 1,000 Neurons with a 10,000/day free allocation W3.
• Security scale: 23.2M network-layer and 29.64T HTTP DDoS requests mitigated in H1 2026; 935 network-layer attacks >1 Tbps P4.
• Analyst recognition: the only vendor named a Visionary in both 2026 Gartner MQ for SASE and SSE P5.
• Content/agentic market data: 2.5B regular genAI users in 3.5 years; open-web time collapsing to 15 minutes per hour of search P7.
• Public interest in the post-quantum post was modest (3 HN points, 0 comments) E1.