Ex-Meta Insider: AI Is About to Change Forever (and nobody even realizes) | Ep 54

If open source models keep getting better, shouldn’t the demand for GPUs collapse? And if your entire business depends on rented access to a single frontier lab, what happens the day that lifeline gets cut?

In this episode of Inside the Silicon Mind, we sit down with Benny Chen, co-founder of Fireworks AI, to dismantle the idea that open source progress destroys GPU demand. Benny spent years working on ASICs and GPU model serving inside Meta, where he was an ads infrastructure lead, before co-founding Fireworks in September 2022 with a team drawn largely from PyTorch and PyTorch-adjacent orgs and near-immediate backing from Benchmark. His team has since tuned models with Cursor for Composer 2 and Composer 2.5, worked with Cognition on their model, and is now working with Harvey on a legal model, with the models Fireworks serves scaling a thousand times over four years. Today, Fireworks focuses on what Benny calls specialized intelligence: helping companies customize open source models, serve them in production, and own the intelligence their products run on instead of renting it.

You’ll hear why Benny believes collaboration is always cheaper than closed research, and why he argues the collective brainpower behind open source will eventually overwhelm even the best-funded closed labs (a timing question, in his view, not a yes-or-no one). He walks through the exact customization playbook Fireworks runs, tuning an open source base model to price competitiveness against frontier models, then building a data flywheel so a business can keep improving that model and swap the base out as better releases land. He also uses the Hugging Face incident to explain why safety filters are structurally asymmetrical, helping attackers more than defenders, and makes the case that if the AI boom unwinds, it could look less like the dot-com bubble and more like a real estate correction, where demand is real and leverage is what takes players down.

Who this episode is for:

• Founders and CEOs whose products currently depend on continued access to a single frontier lab
• CTOs, engineering leaders, and infrastructure teams deciding between open source and proprietary models
• Investors evaluating AI infrastructure businesses and weighing the bubble question
• Security leaders thinking through model access, guardrails, and on-prem deployment
• Engineers and researchers following the open source model ecosystem, from DeepSeek to Kimi

What you’ll learn:

• Why Benny believes a better open source model makes GPUs more valuable, not less, despite what the market’s reaction suggests
• How Fireworks’ customization playbook works: tuning an open base model to price competitiveness, building a data flywheel, and swapping bases as new releases land
• Why the recent DeepSeek release may undercut everything in its quality class on price, and what that signals about where open source is heading
• Why safety filters could be asymmetrical by nature, with attackers routing around guardrails that defenders cannot
• How an AI unwind might resemble a real estate bubble rather than dot-com, and why Benny sees leverage, not demand, as the real risk