01
The TSMC of AI Factories,
F.F Research · AI · 5 November 2026
The TSMC of AI Factories,
Fireworks
Fireworks processes more than 40 trillion tokens a day. It says over 95% of them are custom work. That is a process node, not a GPU marketplace.
02
40T+Daily tokens
>$1BARR
>95%Specialized mix
$17.5BSeries D
Thesis · Company disclosure, July 2026
95% specialized is the foundry test
Specialized>95%
Public weights<5%
Lambert’s reading after Forge: “specialized” means any post-trained customer model. The commercial claim still holds.
03
Article architecture
Team → stack → moat → foundry
- 01
Team and capital
Four priced rounds, then a tender. Seven founders, six Meta, one Google.
- 02
Six modules
Infra through Nexus. Training, Lab, models, inference.
- 03
Moat and why
Skew, async RL, long context. TSMC and vLLM.
04
Chapter 01
Who built the process line
A PyTorch-native bench, priced at $17.5B, that refuses to be a closed lab.
- 1.1 Funding history
- 1.2 Founding team
05
Chapter 1 · 1.1 Funding
Four priced rounds. One tender off the books.
A $25M
Benchmark. Valuation undisclosed.
B $52M
$552M. Sequoia. NVIDIA, AMD.
C $250M
$4.0B. 10k companies, ARR >$280M.
D $1.505B
$17.5B. ARR >$1B. 40T tokens/day.
Sep 2026: $132M Atreides employee tender, cash does not enter the company. Oct 2026 $30–40B talks: unverified.
06
Six Meta, one Google.
Chapter 1 · 1.2 Founders
Six Meta, one Google.
About half from the original PyTorch team.
Lin Qiao
Head of PyTorch. CEO. Fudan, UCSB, IBM, LinkedIn, Meta 300-person org.
Dzhulgakov
PyTorch core maintainer. CTO. Caffe2, ONNX, DLRM.
Benny Chen
Ads infra. GPU then ASIC inference, capacity, TCO.
Chenyu Zhao
The Google hire. Vertex AI, multi-cloud, heterogeneous GPU.
Ivchenko
PyTorch for ranking. TorchRec, large-scale recommenders.
James Reed
Compiler. torch.fx, PiPPy, TorchScript, ONNX.
Garbacki
News Feed core ML. Post-train / RL, Nexus, Ember-1.
07
Chapter 2 · Six modules
Ensembling is the roof. Infra is the floor.
06 Ensembling
Nexus for people. API for systems. Hard turns closed, routine turns open.
05 Inference
Serverless → Reserved → Dedicated as traffic grows.
04 Models
In-house, open library, customer weights. Same stack into production.
03 Lab
Embedded researchers. The SKU is time, not GPU.
02 Training
Managed, Serverless API, Dedicated API, RL rollouts only.
01 Infra
Leased global pool, or the same engine on customer metal.
08
Chapter 3 · Infrastructure
Same engine. Two landings.
Cloud
Virtual pool across 20+ suppliers. Marketing says 30+ regions. Training Available in five. NVIDIA + AMD. Azure MACC via Microsoft Foundry.
vs
Sovereign / Airgap
Managed BYOC on customer GPU. Metadata can still reach the control plane. Airgapped EKS: no console, no overflow, no metadata back.
09
Chapter 4 · Training
Who writes the loop decides the SKU
Managed
Platform writes SFT / DPO / ORPO. FAQ: LoRA only. Pricing still lists full SFT.
Serverless API
Customer writes the loop. LoRA only. Prefill / Sample / Train tokens.
Dedicated API
Full weights, giant MoE. GPU-hours for Trainer + Sampler.
Rollouts only
Customer’s trainer. Hot-load checkpoints on the inference side.
10
Chapter 4 · Public results · marketing numbers
Named lifts, not an audit
| Customer | Method | Published cut |
|---|---|---|
| Harvey Tenet | Kimi K3 + async RL | LAB all-pass 19.7% vs 10.8% |
| Vercel v0 | RFT + spec decode | 93% error-free; ~40× latency vs replaced closed model |
| Heidi Evidence | Nemotron 3 Ultra + RL | ~54% live A/B vs ~42% base |
| Factory | Two Qwen LoRAs | ~70% true leaks at 5% FP vs GPT-5.5 ~59% |
| Kalshi | Glimmer 30B SFT | 58.3% unseen vs base 53.2% |
11
Chapter 5 · Fireworks Lab
High ARPU. The SKU is researcher time.
Diagnosis
Capability, baseline, success criteria, scope. Can we win before more money.
Eval + recipe
Customer experts write the ruler. Base, method, serving economics.
6–10 weeks
Customer keeps model, harness, loop, recipes. GPUs still bill on public SKUs.
12
Chapter 6 · Ember-1 · company self-test
About 40% fewer tokens, quality close to K3
Terminal 2.182.0 vs 80.9
SWE-bench92.2 vs 93.2
DeepSWE 1.175.2 vs 66.4
τ-2 Airline66 vs 64
Two production coding A/Bs: about −35% tokens per task. Closed-model comps are in-house.
13
Chapter 6 · Supply lanes
Three lanes. Weights stay with the customer.
Fireworks models
Ember-1 and in-house specialists. Research Preview, then community demand.
Model library
Day-0 open bases. DeepSeek, Kimi, GLM, MiniMax, Qwen, Step.
Your models
100% customer-owned, base license. Live-merge LoRA. Multi-LoRA ~100 on one Dedicated.
14
Chapter 7 · Inference
Carpool, then a minute of SLA, then the whole card
Serverless
Token carpool. No cold start. OpenAI / Anthropic shapes. Spot and Batch sit here.
Reserved Throughput
Dollars per minute, use-it-or-lose-it. Pricing page: throughput SLA + 99.9% uptime. Docs: throughput only.
Dedicated / Reserve
Whole cards by the second or the year. Private weights, Multi-LoRA, new silicon.
15
Chapter 8 · Nexus
A marketplace finds models. A funnel fills engines.
OpenRouter
Horizontal B2D marketplace. Find, compare, fail over. No fleet. Fireworks is an upstream supplier.
vs
Nexus
Keep Claude Code / Cursor. Keys stay local. Routine turns (~70%) onto Fireworks. $8 / $10 is the wedge.
16
Chapter 9 · Three industrial gaps · company claims
Skew, idle, and context cost
Train–serve skew
MoE experts diverge after FP8/FP4. Router Replay. Stated KLD < 0.007, reject < 3%.
Async RL idle
70–85% of GPU time is rollout. ~2% delta weights, sub-minute hot-swap, 10k+ GPU scale.
Long-context cost
Session Affinity, stated 99.6% KV hit. Nexus dumps routine coding turns. Bill −50–72%.
17
Chapter 9 · Monetization
Bin-pack, then sell the spread, then lock the weights
04 Software
Lab + Sovereign + Nexus Enterprise. Near-zero GPU COGS.
03 Stickiness
Evals, RFT pipes, NVFP4 weights. Harder to swap than a public API.
02 PaaS spread
H100 wholesale ~$2–$3/h. Sold at $8/h after the engine. Source notes, not a filing.
01 Fill the trough
Serverless + Nexus open turns. 10,000+ customers bin-packed into GPU valleys.
18
Chapter 10 · Versus TSMC
IDM sold the brand. Foundry sells the process.
| Semiconductor | Fireworks analog |
|---|---|
| Intel IDM | Closed frontier lab, black-box API |
| Fabless (Apple, NVIDIA) | Cursor, Harvey, Notion, Uber, Vercel |
| EUV tools | H200 / B200 / B300 + MI325X, as yield not rental |
| PDK + standard IP | Open bases + FireOptimizer, Day-0 |
| Never compete with customers | No closed lab. Weights 100% customer-owned |
19
Chapter 10 · Versus vLLM / SGLang
Same single-box problem. Then the factory starts.
| vLLM | SGLang | Fireworks | |
|---|---|---|---|
| Job | Open high-throughput serve | Agents, prefix, JSON | Train-serve factory + online RL |
| Quant | User-tuned | User-checked | Gated KLD < 0.007 |
| Cache | Node prefix | Node radix | Cluster Session Affinity |
| MoE + RL | Inference only | Sampler behind veRL | Router Replay + 2% delta hot-load |