▶ 50:52 · $TOGETHER-AI · Together AI competes on commodity inference price; Fireworks differentiates on customized model quality and zero-KLD training-inference parity

episode briefing

20VC

Are OpenAI & Anthropic Overvalued? The Open-Source AI Reality with Fireworks AI, Lin Qiao

2026-07-20 · 12 company · 8 thematic

sentiment

8 bull2 bear2 neu

speakers

lin qiao

Lin Qiao (Lynn Quo) founded Fireworks AI after 7 years at Meta (Facebook) where she worked on distributed systems and databases. She started the company at age 48 after deliberately building leadership skills. Fireworks provides a specialized intelligence platform for model customization and inference, processing 40T+ tokens/day with $800M ARR.

harry stebbings

Founder of 20VC (The Twenty Minute VC), one of the most popular venture-capital podcasts, and its associated fund.

quote

“Together's price king... if you want cheap you go there... but we're probably not comparing apples to apple... majority of our traffic is customized model... we optimize for quality number one... during training time there's a very hard th…”

—lin qiao

episode shorts · 4

Why AI Independence Matters Now

What Jensen Huang Taught Her About Leadership

Inside AI's Next Big Shift

Why Token Prices Are About to Collapse

now playing · $TOGETHER-AI

$ANTHROPICbearish· highlin qiao

Anthropic's AGI thesis (one model solves all) contradicts specialized intelligence reality; enterprise adoption favors customization

Lin Qiao argues Anthropic's core belief in AGI — a single general model solving all problems best — is fundamentally flawed because every company has unique data, taste, and workf…

$OPENAIneutral· mediumlin qiao

OpenAI and frontier labs build "power line" infrastructure but cannot replace specialized intelligence layer

Frontier model companies (OpenAI, Anthropic, Google) provide essential base infrastructure — like power lines enabling appliances — but the value accrues in specialized intelligen…

$SPCX···bearish· lowlin qiao

SpaceX acquisition of Cursor would create churn risk for Fireworks; industry concentrated on few breakout app companies

Fireworks acknowledges concentration risk: the AI app layer has few escape-velocity companies (Cursor being one), and a SpaceX acquisition of Cursor would threaten Fireworks' reve…

$NVDA···bullish· highlin qiao

Nvidia's Nemotron solves US open-model supply chain risk; Jensen's five-layer cake bottlenecked at energy and chips

Nvidia trains Nemotron open models purely to ensure US supply chain independence for the AI stack's model layer; Jensen Huang's "five-layer AI cake" (application, model, infrastru…

$CURSORbullish· highlin qiao

Cursor grew 1000x in two years on Fireworks; pioneered RL training across distributed data centers for coding agents

Cursor was Fireworks' early coding customer (single-digit millions ARR) and scaled 1000x in two years; they co-developed distributed RL training across 5-6 data center regions wit…

$CRM···bullish· mediumharry stebbings

Salesforce spends 3.8% of developer salaries on Anthropic/Claude Code; signals massive enterprise AI budget allocation

Marc Benioff revealed Salesforce allocates ~3.8% of developer salaries to Anthropic's Claude Code; if this scales to 20% as tools improve, the enterprise AI software market is vas…

$FIREWORKS-AIbullish· high· poslin qiao

Fireworks AI targets 10x token cost reduction in 3 years driving 100x usage growth

Lin Qiao predicts 10x infrastructure cost reduction within three years as supply constraints ease, which will unlock 100x token usage growth; Fireworks' specialized intelligence p…

$TOGETHER-AIneutral· mediumlin qiao

Together AI competes on commodity inference price; Fireworks differentiates on customized model quality and zero-KLD training-inference parity

Together AI is "price king" for off-the-shelf model inference, but Fireworks' majority traffic comes from customized models where quality (zero KLD between training and inference)…

$NVDA···bullish· mediumlin qiao

Nvidia acquired Groq (SRAM accelerator) to combine flops-intensive GPU with SRAM-intensive inference for heterogeneous data centers

Nvidia's acquisition of Groq (SRAM-based ASIC accelerator) enables heterogeneous data center design: flops-intensive GPUs handle prefill while SRAM-intensive accelerators handle t…

$GROQbullish· mediumlin qiao

Groq's SRAM-based accelerator acquired by Nvidia; HBM bottleneck drives heterogeneous inference architecture

Groq's SRAM-intensive ASIC accelerator (acquired by Nvidia) addresses the HBM memory bottleneck for token generation; combined with flops-intensive GPUs for prefill, this heteroge…

$META···bullish· mediumlin qiao

Meta's 5+ year chip investment (MTIA) for ranking/recommendation workloads now extends to AI; workload stability enables hardware specialization

Meta has built custom chips (MTIA) since ~2018 for its massive ranking/recommendation workloads; this long-term investment pays off as workloads stabilize, allowing hardware speci…

$FIREWORKS-AIbullish· high· poslin qiao

Fireworks AI hits $800M ARR, expects to double by year-end on specialized intelligence demand

Fireworks has reached $800M ARR processing 40T tokens/day (majority from customized models) and expects at least 2x revenue growth by year-end; the company's focus on model custom…

←

▶ 50:52 · $TOGETHER-AI · Together AI competes on commodity inference price; Fireworks differentiates on customized model quality and zero-KLD training-inference parity

episode briefing

20VC

Are OpenAI & Anthropic Overvalued? The Open-Source AI Reality with Fireworks AI, Lin Qiao

2026-07-20 · 12 company · 8 thematic

sentiment

8 bull2 bear2 neu

speakers

lin qiao

harry stebbings

Founder of 20VC (The Twenty Minute VC), one of the most popular venture-capital podcasts, and its associated fund.

quote

“Together's price king... if you want cheap you go there... but we're probably not comparing apples to apple... majority of our traffic is customized model... we optimize for quality number one... during training time there's a very hard th…”

—lin qiao

episode shorts · 4

Why AI Independence Matters Now

What Jensen Huang Taught Her About Leadership

Inside AI's Next Big Shift

Why Token Prices Are About to Collapse

now playing · $TOGETHER-AI

$ANTHROPICbearish· highlin qiao

Anthropic's AGI thesis (one model solves all) contradicts specialized intelligence reality; enterprise adoption favors customization

Lin Qiao argues Anthropic's core belief in AGI — a single general model solving all problems best — is fundamentally flawed because every company has unique data, taste, and workf…

$OPENAIneutral· mediumlin qiao

OpenAI and frontier labs build "power line" infrastructure but cannot replace specialized intelligence layer

Frontier model companies (OpenAI, Anthropic, Google) provide essential base infrastructure — like power lines enabling appliances — but the value accrues in specialized intelligen…

$SPCX···bearish· lowlin qiao

SpaceX acquisition of Cursor would create churn risk for Fireworks; industry concentrated on few breakout app companies

Fireworks acknowledges concentration risk: the AI app layer has few escape-velocity companies (Cursor being one), and a SpaceX acquisition of Cursor would threaten Fireworks' reve…

$NVDA···bullish· highlin qiao

Nvidia's Nemotron solves US open-model supply chain risk; Jensen's five-layer cake bottlenecked at energy and chips

Nvidia trains Nemotron open models purely to ensure US supply chain independence for the AI stack's model layer; Jensen Huang's "five-layer AI cake" (application, model, infrastru…

$CURSORbullish· highlin qiao

Cursor grew 1000x in two years on Fireworks; pioneered RL training across distributed data centers for coding agents

Cursor was Fireworks' early coding customer (single-digit millions ARR) and scaled 1000x in two years; they co-developed distributed RL training across 5-6 data center regions wit…

$CRM···bullish· mediumharry stebbings

Salesforce spends 3.8% of developer salaries on Anthropic/Claude Code; signals massive enterprise AI budget allocation

Marc Benioff revealed Salesforce allocates ~3.8% of developer salaries to Anthropic's Claude Code; if this scales to 20% as tools improve, the enterprise AI software market is vas…

$FIREWORKS-AIbullish· high· poslin qiao

Fireworks AI targets 10x token cost reduction in 3 years driving 100x usage growth

Lin Qiao predicts 10x infrastructure cost reduction within three years as supply constraints ease, which will unlock 100x token usage growth; Fireworks' specialized intelligence p…

$TOGETHER-AIneutral· mediumlin qiao

Together AI competes on commodity inference price; Fireworks differentiates on customized model quality and zero-KLD training-inference parity

Together AI is "price king" for off-the-shelf model inference, but Fireworks' majority traffic comes from customized models where quality (zero KLD between training and inference)…

$NVDA···bullish· mediumlin qiao

Nvidia acquired Groq (SRAM accelerator) to combine flops-intensive GPU with SRAM-intensive inference for heterogeneous data centers

Nvidia's acquisition of Groq (SRAM-based ASIC accelerator) enables heterogeneous data center design: flops-intensive GPUs handle prefill while SRAM-intensive accelerators handle t…

$GROQbullish· mediumlin qiao

Groq's SRAM-based accelerator acquired by Nvidia; HBM bottleneck drives heterogeneous inference architecture

Groq's SRAM-intensive ASIC accelerator (acquired by Nvidia) addresses the HBM memory bottleneck for token generation; combined with flops-intensive GPUs for prefill, this heteroge…

$META···bullish· mediumlin qiao

Meta's 5+ year chip investment (MTIA) for ranking/recommendation workloads now extends to AI; workload stability enables hardware specialization

Meta has built custom chips (MTIA) since ~2018 for its massive ranking/recommendation workloads; this long-term investment pays off as workloads stabilize, allowing hardware speci…

$FIREWORKS-AIbullish· high· poslin qiao

Fireworks AI hits $800M ARR, expects to double by year-end on specialized intelligence demand

Fireworks has reached $800M ARR processing 40T tokens/day (majority from customized models) and expects at least 2x revenue growth by year-end; the company's focus on model custom…