Theses›Datacentre & AI compute›AI kernel synthesis

AI kernel synthesis

AI/search-based kernel synthesis attacks the performance-parity-kernel constraint on heterogeneous AI compute, and at expert level on novel silicon it collapses the CUDA library moat

  • Bear
  • Closing
  • $3.1B by 2035

The callRestraints win for now: the tools reach 17-56% of expert speed on new chips, and the incumbents ship their own. An acquisition market, and late.

D1D2R1R2C1foragainst D1D2R1R2C1foragainst
Fig. 1The balance. Drivers and catalysts against restraints, each block sized by its weight; catalysts hang dashed because they have not happened yet, sized by their odds. It tips toward the restraints.
  1. D1AI compute is going multi-vendor, and every new chip needs years of hand-tuned kernels from a few hundred scarce specialists before models run at full speed.
  2. D2Hyperscalers designing their own chips create buyers who cannot hire their way to NVIDIA's library depth.
  3. R1It does not work yet where it matters: the best model scores 56% on PLENA, 17% on the AMD NPU, 47% on Coral, and 46% beat PyTorch on Huawei's chip.
  4. R2The incumbents ship their own tools, strongest on NVIDIA, and every exit (OctoML, Nod.ai, Modular) was an absorption by a chip or platform company.
  5. C1AMD's MI450 Helios racks ship in volume against the OpenAI 6GW agreement, the first gigawatt-scale fleets outside CUDA that need full-speed kernels.end-Q3 2026
Rests onthe beliefs underneath
Supported by · confidence
Threatened by · confidence
What would prove itdated tests, open and resolved
20272028today1234predictionswatch 20272028today1234predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1An independent (non-hyperscaler) kernel-synthesis tool demonstrates expert-level (>=90% of hand-tuned) performance on a NOVEL non-GPU accelerator with little public kernel code, on real silicon (KernelBench-style).Dec 2027 · open
  2. 2An independent tool at 90% or more of expert performance on a new non-GPU accelerator with little public kernel code.Dec 2027 · open
  3. 3At least one of the funded independent kernel-synthesis seeds is acquired by a silicon vendor rather than reaching standalone scale.Jun 2028 · open
  4. 4One of Standard Kernel, Mako, Luminal or Kernelize bought by a chipmaker, confirming the acquisition pattern.Jun 2028 · open
Expressionslisted, private, technologies
  • AMDAMD
    • full
     MI300/MI350 is the only credible non-NVIDIA training silicon at scale.
  • NVDANVIDIA
    • cheap
     Gives away TensorRT compilers and model-optimisation libraries for its GPUs, sells them inside NVIDIA AI Enterprise, and bought OctoAI and CentML.
  • QCOMQualcomm
    • not screened
     Owns Modular (bought Jul 2026) and sells its MAX serving framework, Mojo kernel language and hosted endpoints across chip vendors.
  • IBMIBM (Red Hat)
    • full
     Sells Red Hat AI Inference Server, priced per accelerator, with model compression tools from its Neural Magic acquisition, to enterprises.
  • INTCIntel
    • full
     Distributes the free OpenVINO toolkit for optimising and running models on Intel CPUs, GPUs and NPUs.
Private · 15 private companies, 13 of them in the US; none in continental Europe.

Shinso · Tensormesh · Gimlet Labs · AlphaTango · Standard Kernel · Makora · Luminal · Kernelize · Modular · OctoML · OctoML (OctoAI) · CentML · Neural Magic · Makora (formerly Mako) · Wafer (formerly Herdora)

Technologies
Who has already bet on itprivate rounds, from public sources

$667M raised by 9 companies, $324M of it in the last three years. Wave under way: Large rounds in the last three years; the category is funded and getting crowded.

Backed by a top-tier venture firm: Modular, OctoML (OctoAI), Neural Magic, Luminal. Backed by a strategic investor: CentML.

1 more funded company in the market.

What changedthe record, newest first
  • 6 Jul 2026Saw the algorithm-to-hardware synthesis pitch again this week, this time pointed at defence control loops instead of datacentre kernels. Same conclusion: the clever compiler is real, but whoever owns the silicon ends up owning the compiler.
  • 3 Jul 2026The neutral compiler layer is too valuable to stay neutral, the independent gets bought by silicon (OctoML, now Modular at ~$3.9bn). Being bought is not the same as failing to capture value.
  • 23 Jun 2026The first benchmark for AI-generated kernels on brand-new chip architectures came in well below parity (17-56%), while the same tools keep getting better on mature GPUs. The capability is strongest exactly where the incumbents are already strong, which is the opposite of what would crack the moat.
Reading
Show the reading