Fireworks CEO乔琳：有PMF也可能破产，AI不是SaaS

Summary

• PMF in AI doesn’t guarantee scalability, and rapid growth can still drain your cash. 曹卿云 says AI infrastructure is expensive; a small company that grows too fast without securing its next round could “scale into bankruptcy.” A large company rolling out an AI feature to all users could also face eye-watering costs and get blocked by its CFO.
• The economic value of specialized intelligence lies in lower costs and supply-chain control. 曹卿云 says customers can cut costs by 5–10x. Closed-source vendors control upgrades and capacity, and may be both vendor and competitor. Building on open-source models and customizing them with proprietary data is one effective way to control the supply chain.
• The Specialized Intelligence Index uses companies’ private benchmarks to measure model performance in specific domains. 曹卿云 says public benchmarks are saturated, while companies’ unique products and tasks require equally unique evaluations. The Index will track the quality of new open- and closed-source models and how they advance domain-specific tasks.
• Amber One cuts costs by shortening chains of thought; the platform jointly optimizes training and inference. 曹卿云 says chains of thought can be cut in half, halving token usage and cost. He also says customized models can deliver bitwise-equivalent inference with no loss of accuracy.
• Applications are expanding beyond coding into knowledge work, while consumer-scale adoption remains conditional. 曹卿云 cites healthcare, life sciences, customer support, recruiting, marketing, sales, and cybersecurity. Consumer applications are still early; this year’s experiments must validate user demand, PMF, and scalability. He sees a strong chance of widespread consumer applications next year.
• The challenge over the next 18 months is keeping pace with superlinear demand without losing speed. 曹卿云 is focused on preserving efficiency and velocity as the company grows, through a lean, cross-functional structure with end-to-end ownership.

Deep dive

1. Specialized intelligence needs to be measured against companies’ own tasks

• 曹卿云 sees specialized intelligence and general intelligence as parallel concepts. Most of the world’s data has not been absorbed by foundation models because it remains inside applications, products, and companies rather than being shared externally. The frontier of intelligence is therefore also inside those applications and companies, which should turn their private data into proprietary agents.
• He explains that a company’s distinctiveness comes from solving a unique problem in a unique way; a product’s distinctiveness is reflected in its evaluation. Public benchmarks are saturated, so high scores after training on benchmark data don’t tell us much. The Specialized Intelligence Index will aggregate companies’ private benchmarks and track the quality of new open- and closed-source models, helping the industry assess how they advance tasks in each domain.
• The host mentions the Index and a model released on the 23rd. 曹卿云 says their release dates were close, but there was no direct connection between them.

2. After PMF come two more hurdles: costs and the supply chain

• 曹卿云 draws a distinction between SaaS and AI. In the SaaS era, PMF and a durable business were essentially the same thing, after which a company could scale quickly. Labor, storage, and CPU were the main costs, and storage and CPU had become commoditized. AI infrastructure, by contrast, is expensive, so PMF in AI doesn’t guarantee a company can scale; a small company that grows quickly but can’t raise its next round could “scale into bankruptcy.”
• Large companies face the same problem. A company with a large base of consumer or developer users that rolls out an AI feature to everyone could face eye-watering costs, leading its CFO to block the launch. 曹卿云 says specialization can cut customers’ costs by 5–10x, with a major impact on the business.
• Closed-source model vendors control model upgrades and the capacity they can provide. If a customer needs to grow 10x and its vendor can’t support that, the customer hits a bottleneck. The vendor may also launch its own product, becoming both vendor and competitor. 曹卿云 says using open-source models and customizing them with a company’s own data is an effective way to control the supply chain.

3. Cutting model costs without compromising inference quality or service

• Amber One was released to make models smarter while building on open-source models to lower costs. 曹卿云 says it can cut a model’s chain of thought in half, halving token usage and cost. When the host says the results are the same, 曹卿云 agrees.
• As a platform company, Fireworks also takes customers’ shared needs and addresses them at the model layer. That includes extending the context window to support long-horizon agents with hundreds or even thousands of turns, and making models more concise so they can solve problems faster—saving both money and time. He says Amber One iterates on other open-source models and believes the open-source community can have the greatest impact by building on one another’s work.
• On the competitive landscape, 曹卿云 stresses that training and inference are co-designed and co-optimized. Once a model has been customized, inference can be bitwise equivalent, with no loss of accuracy. A long-horizon agent might take 100 turns to complete a task; if each turn has 0.1% divergence, that could accumulate to 10%, making the result unusable. He says Fireworks can eliminate divergence and optimize inference during training.
• Their end-to-end platform also includes Nexus smart router. Customers can combine closed-source, off-the-shelf open-source, and customized models without having to decide when to use each one; the router automatically sends each request to the right model based on customer needs.

4. Knowledge-work assistants are expanding; consumer adoption still needs validation

• 曹卿云 says coding was the only application category to take off last year. This year, coding is largely solved, and knowledge-work assistants are spreading across fields: physician or patient assistants, drug discovery, cancer treatment, applications to ease clinical-trial shortages, customer support, recruiting, marketing, sales, and cybersecurity. He singles out efforts to automate the process from detection through defense in cybersecurity.
• He says consumer applications are still at an early stage. This year’s experiments need to establish whether products meet user needs and solve real problems, and whether they can scale into a business once they have PMF. Some applications are already gaining traction, so he sees a strong chance of consumer applications becoming widespread next year—but that remains a forecast, not a certainty.

5. Scaling the organization to meet superlinear demand without slowing down

• Looking ahead to the next 18 months, 曹卿云 is most focused on scaling the company to meet superlinear demand without losing velocity. For him, that is both a question of growth speed and organizational design.
• Fireworks has few layers, and no single team can deliver on its goals alone; teams need end-to-end ownership of cross-functional work. He is rethinking the organization from first principles: as the company gets bigger, how can it preserve that collaborative efficiency and scale as quickly as possible without slowing down?