ValuationsMarks, multiples & secondaries
Fireworks AI co-founder says most companies can skip frontier models, putting open models three to six months behind
At Slush, Fireworks AI co-founder Chenyu Zhao said open models trail frontier systems by three to six months. He urged companies spending $10,000–$100,000 a month on tokens to consider switching. Fireworks disclosed two breaches this autumn that affected customers.
From the TechCrunch · Venture article of November 20, 2026 and cited sources: Zhao expects a 'two-model world.' Frontier models would handle high-level planning, and cheaper customized open models would run routine steps. [1]
SOURCES[1] Anthropic rival says companies can (mostly) skip the… techcrunch.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Zhao expects a 'two-model world.' Frontier models would handle high-level planning, and cheaper customized open models would run routine steps. [1]
- Original source
- TechCrunch · Venture article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 20, 2026 article published
- Covered
- Nov 20 edition
- Evidence
- 15 sources across 12 sites
- 1Anthropic rival says companies can (mostly) skip the…techcrunch.com
- 2BleepingComputer: Uber notifies driversbleepingcomputer.com
- 3CNBC: Anthropic IPO raises over $81 billioncnbc.com
- 4Fireworks: Series D announcementfireworks.ai
- 5CNBC: Fireworks hits $17.5B valuationcnbc.com
- 6Fortune: Nadella warns about know-howfortune.com
Reach: not measured yet

Zhao expects a 'two-model world.' Frontier models would handle high-level planning, and cheaper customized open models would run routine steps. He argues buyers should track cost per task rather than cost per token. [1]
Fireworks raised $1.505B at a $17.5B valuation in July. Atreides, Index Ventures and TCV led the round, and Nvidia participated. The company also says it has passed $1B in annualized revenue, a self-reported figure. [4] [5]
Two October intrusions exposed customer API keys, billing records, and for some accounts, datasets and partial model weights. Downstream, 149,346 Uber drivers were notified, and Notion and Cursor reported exposure. Notion's chat transcripts leaked through a batch feature that stores data at rest, even though Notion had a zero-retention commitment. [2] [7] [8] [9] [10]
Evidence shows heavy capital flowing into open-model customization, including a $17.5B valuation and a $4.5B acquisition.
Read the full assessment
The implication, cheaper AI that protects a company's know-how, rests on untested capability claims, and the breaches show retention guarantees can vary by feature.
Executive brief
Fireworks AI raised money at a $17.5 billion valuation in July. Its pitch is that most companies don't need frontier models. That pitch now comes from a company that disclosed two breaches this autumn, and one of them exposed data on 149,346 Uber drivers (BleepingComputer). Speaking at Slush on the day Anthropic began trading after a record $81.3 billion IPO (CNBC), co-founder Chenyu Zhao said open models trail frontier models by three to six months. He argued that companies spending $10,000–$100,000 a month on tokens should consider switching (TechCrunch).
What changed and event timeline
Nadella's "pay twice" warning
Microsoft's CEO argued that companies pay AI labs once in money and again in leaked institutional know-how. He called for tuning inside the customer's own tenant ().
Fireworks Series D
Fireworks announced $1.505B at a $17.5B valuation, led by Atreides, Index Ventures and TCV, with Nvidia participating. The company says it has passed $1B in annualized revenue (;).
First Fireworks incident
Attackers misused internal credentials to read customer API keys stored as secrets in evaluation jobs. Fireworks linked the group to "LLMjacking", meaning stealing model access to resell tokens ().
Second breach via cloud provider
Attackers went from a compromised third-party cloud host into Fireworks development machines. They reached billing records and, for some accounts, secrets, datasets and incomplete model weights ().
Downstream disclosures
Uber notified 149,346 drivers. Notion disclosed that chat transcripts were exposed and disconnected from Fireworks. Cursor also confirmed it was affected (;).
Applied Compute deal and Anthropic IPO
Palo Alto Networks agreed to buy Applied Compute, another company that customizes open-weight models, for $4.5B in stock. Anthropic began trading the next day (;).
Capabilities and access
- Fireworks hosts open-weight models, fine-tunes them on customer data and serves them through an API.
- Fireworks says it serves more than 40 trillion tokens a day, and that over 95% of them come from models specialized on customer data (Fireworks blog).
- The core inference product is zero data retention (ZDR) by default. Batch prediction, evaluation and training features store data at rest (CSA).
Read the full section
- Fireworks hosts open-weight models, fine-tunes them on customer data and serves them through an API. None of the reviewed sources names specific model versions; Zhao only mentions partnering on Nvidia's Nemotron family (TechCrunch).
- Fireworks says it serves more than 40 trillion tokens a day, and that over 95% of them come from models specialized on customer data (Fireworks blog).
- The core inference product is zero data retention (ZDR) by default. Batch prediction, evaluation and training features store data at rest (CSA).
Technical analysis for researchers and developers
- Routing: Zhao expects a "two-model world": frontier models for high-level planning, cheaper open models for routine steps.
- Retention depends on the feature: Notion's data leaked through batch prediction, which keeps uploaded datasets at rest.
- Secrets in jobs: Keys stored as environment variables in evaluation jobs were a direct path for attackers (BleepingComputer).
Read the full section
- Routing: Zhao expects a "two-model world": frontier models for high-level planning, cheaper open models for routine steps. He says the metric to watch is cost per task, not cost per token. No benchmark or task-level evaluation supports this (TechCrunch).
- Retention depends on the feature: Notion's data leaked through batch prediction, which keeps uploaded datasets at rest. Fireworks says it is moving those services to an online flow that keeps data only as long as needed.
- Secrets in jobs: Keys stored as environment variables in evaluation jobs were a direct path for attackers (BleepingComputer).
Claims and evidence
- Round size, valuation and leads: announced by the company and matched by CNBC's independent reporting (CNBC). Neither source cites a filing.
- $1B+ annualized revenue and 5x growth: self-reported.
- "Five to 10 times cheaper" than equivalent closed models: an estimate from CEO Lin Qiao, not independently tested (CNBC).
Read the full section
- Round size, valuation and leads: announced by the company and matched by CNBC's independent reporting (CNBC). Neither source cites a filing.
- $1B+ annualized revenue and 5x growth: self-reported.
- "Five to 10 times cheaper" than equivalent closed models: an estimate from CEO Lin Qiao, not independently tested (CNBC).
- Open models only three to six months behind: Zhao's opinion; the reviewed sources contain no supporting evaluation.
- New funding talks: reported without named sources; Zhao declined to comment. Not corroborated.
- Applied Compute price: confirmed in a filing (SEC 8-K).
Context and prior work
- Fireworks is part of a funded category of platforms for customizing and serving open models.
- Palo Alto expects the Applied Compute deal to add more than $1B in annual recurring revenue (ARR) by fiscal 2029, a company projection (SiliconANGLE).
- Nadella's "reverse information paradox" essay gives large companies an IP argument for open models, separate from cost (The New Stack).
Read the full section
- Fireworks is part of a funded category of platforms for customizing and serving open models. TechCrunch names Together AI and Baseten as rivals with large recent rounds, but the reviewed sources do not give those round sizes.
- Palo Alto expects the Applied Compute deal to add more than $1B in annual recurring revenue (ARR) by fiscal 2029, a company projection (SiliconANGLE).
- Nadella's "reverse information paradox" essay gives large companies an IP argument for open models, separate from cost (The New Stack).
Limitations, safety and contested findings
- Zhao says there is no sign that zero retention for inference traffic was compromised.
- Fraudulent model calls made with stolen keys have run up cloud AI bills as high as $1M, according to secondary security coverage (Aviatrix).
- Fireworks has not given a total count of affected customers. Cursor reported backdoor malware on users' devices (PYMNTS).
Read the full section
- Zhao says there is no sign that zero retention for inference traffic was compromised. The Cloud Security Alliance (CSA) note finds that Notion's contractual ZDR commitment did not stop transcripts from being stored through a secondary feature (CSA).
- Fraudulent model calls made with stolen keys have run up cloud AI bills as high as $1M, according to secondary security coverage (Aviatrix).
- Fireworks has not given a total count of affected customers. Cursor reported backdoor malware on users' devices (PYMNTS).
Business and practitioner implications
- When to switch: Zhao's threshold of 10,000–100,000 users or $10,000–$100,000 a month in token spend is a heuristic from a vendor.
- Trade-off: you gain cost savings and keep your know-how, but you take on subprocessor risk.
- Consolidation: security vendors are now buying model-customization startups, as Palo Alto's deal shows.
Read the full section
- When to switch: Zhao's threshold of 10,000–100,000 users or $10,000–$100,000 a month in token spend is a heuristic from a vendor.
- Trade-off: you gain cost savings and keep your know-how, but you take on subprocessor risk. Check ZDR endpoint by endpoint, keep secrets out of job environments, and rotate any keys issued before late October 2026 (CSA).
- Consolidation: security vendors are now buying model-customization startups, as Palo Alto's deal shows.
The source trail.
Sources (15)
Anthropic rival says companies can (mostly) skip the pricey models
Article text retrieved; extracted text may omit tables or interactive elements.
techcrunch.com