CHAPTERS
- 0:05 – 0:58
Wafer’s core product: a fast AI inference cloud optimized by agents
Jared kicks off by framing Wafer’s rapid rise, and Emilio defines the company: an inference cloud that focuses on best-in-market speed. The differentiator is using AI agents to optimize GPU performance end-to-end.
- •Wafer is an AI inference cloud focused on latency and throughput
- •Core method: agents that optimize GPUs (and the software stack)
- •Goal: best speeds in the market as the primary wedge
- 0:58 – 1:36
Who uses Wafer: voice agents and high-throughput coding workloads
Emilio explains the two major usage patterns Wafer serves: low-latency real-time products and high-throughput token generation. He highlights voice agents (where seconds of delay ruin UX) and coding agent workloads (where throughput is paramount).
- •Voice agents need ultra-low latency to feel responsive
- •High-throughput workloads (e.g., coding agents) benefit from maximum tokens/sec
- •Example customer type: platforms like Vercel optimizing coding agent speed
- 1:36 – 1:53
The $40M Series A and why investors leaned in
The team announces a $40M Series A and shares who led it. The raise is positioned as fuel to scale rapidly amid surging demand and compute constraints.
- •$40M Series A co-led by Marathon and Chemistry
- •Strong angel participation; existing investors include Jeff Dean and YC
- •Capital supports scaling infrastructure to meet demand
- 1:53 – 3:11
Zero to $8M ARR in four months: the GLM 5.2 moment and demand shock
Emilio breaks down the inflection point: shifting strategy to use their optimization agent internally on open-source LLMs and offering the fastest versions broadly. Their GLM 5.2 optimization delivered 2–3x performance, triggering a spike in demand that quickly outstripped available GPU capacity.
- •Strategic shift: stop selling the optimizer; use it internally to serve optimized open-source models
- •GLM 5.2 launch: rewrote kernels and optimized the full GPU stack
- •Delivered 2–3x performance vs market, driving sudden adoption
- •Growth forced a funding round to keep up with GPU needs
- 3:11 – 5:29
When launch success breaks operations: scrambling for GPUs and onboarding chaos
Jared recounts pre-launch office hours advice—launch first, then raise once demand proves out—which played out exactly. Steven describes the operational reality: onboarding new nodes, debugging reliability and networking, and keeping the system running while demand surged.
- •YC office hours: bet on a strong launch to create an investor narrative
- •Post-launch: compute became the bottleneck immediately
- •Steven’s role: securing GPUs, onboarding nodes, debugging reliability/networking
- •Team intensity: even offsites turned into infrastructure sprints
- 5:29 – 6:18
Customer case study: Neon Health and the latency-driven switch
Emilio explains how Neon Health uses AI agents to automate hospital phone workflows and why latency is business-critical. Wafer’s 30–50% performance improvement per call justified switching from a larger inference provider.
- •Neon Health automates hospital processes with AI voice agents
- •Latency directly impacts completion rates and user experience
- •Wafer delivered ~30–50% better per-call performance
- •Demonstrates why speed wins even against larger incumbents
- 6:18 – 8:18
YC’s Office Hour Simulator: a real-time benchmark for ultra-low latency
Jared describes YC’s own product—a video-and-voice “AI clone of a YC partner”—where milliseconds matter for realism. He shares that Wafer won a head-to-head latency benchmark across providers, enabling YC to switch to Wafer’s fast GLM 5.2 setup.
- •Office Hour Simulator requires fast, natural conversational timing
- •Benchmarking focused on lowest latency while maintaining intelligence
- •Wafer outperformed competing providers in clean comparisons
- •Switch from Gemini/OpenAI minis to Wafer-optimized GLM 5.2
- 8:18 – 10:03
A/B test results: faster responses changed user behavior
Jared reveals an unreleased metric: a three-way A/B test showed not only lower latency but also longer user engagement with Wafer’s model. The discussion connects speed to product design—reducing the need for workaround “scaffolding” when the model can respond live.
- •Three-way A/B test: Wafer vs two proprietary models
- •Users stayed significantly longer when latency decreased
- •Speed enables more “live” behavior and reduces surrounding scaffolding
- •Reinforces “speed is the moat” in real-time experiences
- 10:03 – 11:21
How it works: “AI that optimizes AI” across kernels, decoding, quantization, and hardware
Steven details Wafer’s technical approach: take a customer’s workload characteristics and feed them into an AI-driven optimization pipeline. The system can produce custom kernels, train speculative decoding models, quantize models, and even bring up alternative hardware to deliver large speedups.
- •Input: workload characteristics from each customer/use case
- •Agent outputs: custom kernels, speculative decoding models, quantization, runtime tuning
- •Result: per-workload hyper-optimized engines delivering 2–3x+ speedups
- •Extends beyond NVIDIA: exploring AMD for perf-per-dollar parity
- 11:21 – 12:51
From “cursor for CUDA” to cloud: the long technical runway behind the overnight success
The conversation rewinds to Wafer’s origins in YC: they began as an agent-based tool to optimize CUDA kernels and gradually moved up the stack. Prior contracts (e.g., AWS, DigitalOcean) built the team’s optimization muscle, making the later shift to an inference cloud comparatively straightforward.
- •Early concept: “cursor for CUDA” using agents to optimize kernels
- •Progression: moved up the ML performance stack over time
- •Did optimization work for major infrastructure companies (e.g., AWS, DigitalOcean)
- •The cloud pivot leveraged capabilities developed over months
- 12:51 – 14:22
The night the agent beat NVIDIA libraries: deciding to verticalize
Emilio recounts a pivotal demo prep where Steven ran the agent overnight and achieved optimizations surpassing core NVIDIA libraries. That result triggered a strategic rethink: rather than sell optimizations to inference providers, Wafer would compete directly by offering the optimized models themselves.
- •Overnight agent run produced better-than-NVIDIA-library performance in key areas
- •Realization: the optimization value might be too strategic to “give away”
- •Pivot decision: verticalize into an inference cloud instead of selling tooling
- •Set the stage for the later GLM-driven breakout
- 14:22 – 18:14
Open-source LLM adoption goes exponential—and speed (not just cost) becomes the killer feature
They discuss why open-source adoption is accelerating: enterprises see large cost savings, but also new product possibilities from dramatically faster responses. Jared emphasizes a key insight: for latency-sensitive products, open-source models served on Wafer can be the premium option even when price is irrelevant.
- •Large companies are increasingly inbounding to adopt open source
- •Cost savings matter (e.g., big reductions vs proprietary tokens)
- •But speed is the stronger wedge for latency-sensitive products
- •Open-source + Wafer can outperform proprietary options even at higher willingness to pay
- 18:14 – 21:42
Origin story at UChicago: Yakka as a “build for fun” project that became Wafer
Jared takes the story back to how the founders met and the precursor project, Yakka, which iteratively used an LLM to optimize code and test improvements. The project reframed AI as a compiler-like system that can better map software to hardware—an idea compelling enough to turn into a long-term mission.
- •Yakka: loop of “optimize code → measure → repeat” using an LLM
- •Insight: AI can act like a new kind of compiler for hardware mapping
- •Motivation: wasted energy/compute from poorly optimized software
- •Founders chose a problem they could work on for decades
- 21:42 – 23:32
Turning down jobs and “you don’t need a business co-founder” as the activation energy
They describe the mindset shift from taking prestigious jobs to starting a company. Emilio credits a YC talk that removed the perceived requirement for a business co-founder, while Steven emphasizes ownership and pursuing high-impact work as key motivators.
- •Both founders initially had strong job offers (Two Sigma, Series A startup)
- •YC advice: engineering founders don’t need a business co-founder
- •Decision drivers: fun of building together, ownership, societal value
- •Commitment: apply to YC, move to SF, start Wafer
- 23:32 – 26:41
How to pick a rocket ship, hiring at Wafer, and what “Member of Technical Staff” does
Emilio shares career advice: evaluate whether a company can 1000x rather than fixate on a static role, because responsibilities change quickly in fast-growing startups. They close with Wafer’s hiring philosophy—seeking exceptional problem-solvers—and Steven outlines the day-to-day as variable, high-leverage ML systems problem solving.
- •Advice: join companies with true 1000x potential; roles will evolve rapidly
- •Evaluate founders/culture more than job description
- •Hiring: “Member of Technical Staff” across the board; GPU experience not required
- •Day-to-day: solve diverse ML performance/systems problems (kernels, reliability, infrastructure)
