Baud GTM brain

Baud is developing AI accelerator chips and early cluster access for organizations training and serving models.

category

AI training and inference accelerators

market profile

  • text: Frontier model developers face high compute costs and depend on accelerator capacity for training and inference.
  • text: Nvidia and custom silicon vendors compete across accelerator systems and hosted inference, with different maturity and workload coverage.

personas

  • title: Head of AI infrastructure
    pains:
    • Insufficient accelerator capacity for training and serving
    • Evaluating performance per dollar and power under real workloads
  • title: Machine learning engineering lead
    pains:
    • Expensive pretraining, fine-tuning and inference
    • Porting PyTorch models to new hardware without rewriting kernels
  • title: CTO of an AI model company
    pains:
    • Compute costs restrict model experimentation and deployment
    • Early hardware risk and software compatibility complicate procurement

icp

  • text: AI labs and model-driven companies with material pretraining, fine-tuning, reinforcement learning or inference workloads.
  • text: Engineering teams using PyTorch and open model architectures that can evaluate an early access accelerator cluster.

message pillars

  • text: Reduce the compute barrier to building frontier models.
  • text: Build specialized training and inference hardware around a new arithmetic representation.
  • text: Preserve access to existing PyTorch workflows.
  • text: Invite early users to shape and evaluate the cluster.

objections

  • objection: Are the performance claims proven on production silicon?
    response: Request workload-specific benchmark evidence and hardware availability; YC describes an FPGA emulation demo.
  • objection: Will our existing models run without extensive rewrites?
    response: Test representative PyTorch exports and numerical accuracy with the compiler rather than assuming universal compatibility.
  • objection: When can we get capacity and at what price?
    response: Apply for early access and confirm availability, service terms and pricing directly.

competitors

  • name: NVIDIA
    why: Dominant supplier of GPUs for AI model training and inference.
  • name: Cerebras
    why: Custom accelerator systems target both model training and inference.
  • name: SambaNova
    why: Custom AI accelerator systems and cloud offerings compete for model workloads.

keywords

  • AI training accelerator
  • AI inference chips
  • PyTorch accelerator hardware
  • GPU alternatives for AI training
  • AI compute cost reduction
  • custom AI ASIC

voice

  • Technical
  • Ambitious
  • Efficiency-focused

signals

  • text: A team is scaling pretraining or inference and reports GPU capacity or compute cost constraints.
  • text: A model infrastructure team is testing hardware alternatives while retaining PyTorch workflows.

stack

  • PyTorch
  • Qwen
  • DeepSeek
  • GLM
  • Flux
  • Wan

pricing

model: unknown

motion

sales_led

channels

  • early access applications
  • YC directory
  • LinkedIn

confidence

medium

Founders

  • name: Sarang Zambare
    title: Founder/CEO
    role: CEO
    bio: Multi-time founder with 4 patents and 7+ years in deep learning and AI hardware. ML lead for Peloton Guide from concept to 100,000+ devices shipped. Previously founding ML engineer at Peloton and Caper (acquired by Instacart). Also a certified pilot.
  • name: Eric Taylor
    title: Founder & Chief Hardware Architect
    role: Founder
    bio: Experienced ASIC Architect with over a decade of industry experience, including four tape-outs, two major IP releases, two patents, and a publication. Prior roles at NVIDIA, Freescale/NXP, Arteris IP, and Enfabrica.

Back to YC companies