Is Fireworks AI profitable?

No public profitability figures are disclosed. Fireworks reported surpassing $1B in annualized revenue run rate by July 16, 2026, a 5x year over year increase, but infrastructure serving margins in this space typically run 45 to 65 percent gross, well below SaaS norms, and no net income data has been released.

How much funding has Fireworks AI raised?

Fireworks closed a $1.505B Series D on July 16, 2026 at a $17.5B valuation, led by Atreides Management, Index Ventures and TCV. Total funding raised across seven rounds is roughly $1.8B, up from a $4B valuation just nine months earlier in October 2025's $250M Series C.

Who are Fireworks AI's biggest competitors?

Together AI (raised $800M at $8.3B valuation, July 2026) and Baseten (raised ~$1.5B at ~$13B valuation, June 2026) are the closest peer rivals. Hyperscalers AWS, Azure and Google Cloud, plus silicon players Groq (NVIDIA-acquired ~$20B, Dec 2025) and Cerebras (IPO'd ~$66B, May 2026), also compete.

Could hyperscalers replace Fireworks AI?

It's a live risk: AWS Bedrock, Azure AI and Google Vertex already host open-weight models inside existing cloud contracts, which could commoditize Fireworks' serving layer. Fireworks' defense is workflow lock-in via fine-tuned models, claiming 95 percent of served tokens now come from customized models as of July 16, 2026.

What is Fireworks AI's moat against competitors?

Fireworks points to fine-tuned model lock-in (95 percent of tokens from customized models, per July 2026 claims) and proprietary kernel IP like FireAttention and FireCompile. Its Training platform entered preview only in April 2026, so the lock-in advantage is unproven at scale.

Has Fireworks AI shipped a training product yet?

Fireworks Training launched in preview on April 6, 2026, supporting SFT, DPO, RFT and LoRA techniques and powering RL rollout infrastructure behind Cursor's Composer 2 model. It has not yet reached general availability or published gross retention data on customized workloads.

AI companies and labs · United States

Fireworks AI

fireworks.ai ↗

Fireworks AI runs a managed fast inference platform for open source and custom LLMs.

Losing its protection

Assessed 20 July 2026 · Evidence as of 20 July 2026

The structural protection is eroding.

On the recordSince 20 July 2026 · Checkable from around October 2026 · Every call on the record Checkpoint: Fireworks Training platform expected to move from preview to general availability

The call

Fireworks has real revenue and a lock-in story still in preview, sitting between hyperscalers on one side and better-funded inference rivals on the other.

Fireworks AI rents computing power to companies that want to run open-source AI models quickly and cheaply, rather than using big cloud providers like Amazon or Google. The company claims its real advantage is that customers train custom models on Fireworks and would be locked in to keep using it, but this lock-in feature just launched in preview three months ago and has not been proven to actually keep customers around. Right now, Fireworks faces three hard problems at once: rival companies like Together AI and Baseten are winning customers today, Amazon and Google are adding free open-model serving to their existing services to undercut Fireworks' prices, and the profit margins in infrastructure rental are thin (45 to 65 percent) compared to typical software businesses (80 to 90 percent), making it hard to invest faster than competitors. The one thing that would change everything is if Fireworks' training tool becomes fully available and clearly keeps customers from leaving, proving that the lock-in strategy actually works.

Why

AI pressure: AWS Bedrock, Azure AI, Google Vertex managed fine-tuning. If hyperscalers bundle fast custom-model serving into existing cloud contracts, they can undercut Fireworks on price and convenience. Could AI do the core job? Unlikely.

Bigger players: NVIDIA-Groq and Cerebras. NVIDIA now owns Groq and invests in Fireworks at the same time, while Cerebras chases the same latency-sensitive workloads after its IPO. Could a giant copy it? Yes, easily.

What protects it. Customer models fine-tuned on Fireworks now generate 95% of tokens served, creating switching costs if that retention holds up.

What would change this call

Fireworks Training platform ships to GA (targeted ~early 2026) with published retention data. Would show fine-tuned-model lock-in is real, not just a preview feature.
Together AI or Baseten announces a major logo loss or a down round within the next year. Would show peer rivals are losing ground in the same inference market.
AWS, Azure, or Google announces native managed fine-tuning/serving for open-weight models at lower price. Would show hyperscalers are commoditizing Fireworks' core offering.

Key numbers

$1B+ ARR, 5x year over year (Jul 2026)
$17.5B valuation, Series D (Jul 16, 2026)
40T+ tokens served per day (Jul 2026)

Named risks

Thin infrastructure gross margins (45-65%) versus 80-90% for typical SaaS compress profitability. Severity: high
Hyperscalers bundling managed fine-tuning into Bedrock/Vertex could erode Fireworks' differentiation. Severity: high
Fireworks Training platform still in preview as of April 2026, lock-in advantage unproven at GA. Severity: medium
Unaudited company-reported revenue and token metrics have not been independently verified. Severity: medium
Three-way peer fight with Together AI and Baseten could trigger price wars. Severity: medium

Rivals, as read in July 2026

Rival	Direction	Evidence
Together AI	contesting	Raised $800M at $8.3B valuation with $1.15B+ bookings, July 2026
Baseten	contesting	Raised ~$1.5B at ~$13B valuation, ARR estimated ~$600M, June 2026
AWS Bedrock, Azure AI, Google Vertex	losing to	Hyperscalers bundle open-weight model serving into existing cloud contracts
NVIDIA-Groq	gaining on	NVIDIA acquired/licensed Groq for ~$20B in December 2025, deepening hardware rivalry
Cerebras	gaining on	IPO'd May 2026 at ~$66B market cap, competing on latency-sensitive workloads

Questions people ask

Is Fireworks AI profitable?: No public profitability figures are disclosed. Fireworks reported surpassing $1B in annualized revenue run rate by July 16, 2026, a 5x year over year increase, but infrastructure serving margins in this space typically run 45 to 65 percent gross, well below SaaS norms, and no net income data has been released.
How much funding has Fireworks AI raised?: Fireworks closed a $1.505B Series D on July 16, 2026 at a $17.5B valuation, led by Atreides Management, Index Ventures and TCV. Total funding raised across seven rounds is roughly $1.8B, up from a $4B valuation just nine months earlier in October 2025's $250M Series C.
Who are Fireworks AI's biggest competitors?: Together AI (raised $800M at $8.3B valuation, July 2026) and Baseten (raised ~$1.5B at ~$13B valuation, June 2026) are the closest peer rivals. Hyperscalers AWS, Azure and Google Cloud, plus silicon players Groq (NVIDIA-acquired ~$20B, Dec 2025) and Cerebras (IPO'd ~$66B, May 2026), also compete.
Could hyperscalers replace Fireworks AI?: It's a live risk: AWS Bedrock, Azure AI and Google Vertex already host open-weight models inside existing cloud contracts, which could commoditize Fireworks' serving layer. Fireworks' defense is workflow lock-in via fine-tuned models, claiming 95 percent of served tokens now come from customized models as of July 16, 2026.
What is Fireworks AI's moat against competitors?: Fireworks points to fine-tuned model lock-in (95 percent of tokens from customized models, per July 2026 claims) and proprietary kernel IP like FireAttention and FireCompile. Its Training platform entered preview only in April 2026, so the lock-in advantage is unproven at scale.
Has Fireworks AI shipped a training product yet?: Fireworks Training launched in preview on April 6, 2026, supporting SFT, DPO, RFT and LoRA techniques and powering RL rollout infrastructure behind Cursor's Composer 2 model. It has not yet reached general availability or published gross retention data on customized workloads.

Timeline

2026-07-16Closes $1.505B Series D at $17.5B valuation, surpasses $1B ARR run rate
2026-07-15President George Hu (ex-Twilio COO) confirmed leading GTM scaling
2026-06-22Together AI raises $800M at $8.3B valuation, intensifying peer competition
2026-04-06Fireworks Training platform launches in preview, powers Cursor Composer 2
2025-11-10Reinforcement Fine-Tuning (RFT) expands after June 2025 beta
2025-10-28Closes $250M Series C at $4B valuation, reports $280M+ ARR
2024-07Closes $52M Series B at $552M valuation led by Sequoia, NVIDIA, AMD
2024-01Closes $25M Series A led by Benchmark
2022-10Founded by Lin Qiao and six co-founders, most ex-Meta PyTorch engineers

Revision history

20 July 2026

First assessed. Losing its protection. The text above is fixed as of this date and is not edited afterwards.

In the same band

Safe Superintelligence Inc.Mistral AI Cohere