Skip to main content

AI Chip Startup Survival Report: Tenstorrent / SambaNova / Graphcore in 2026

· 8 min read
Industry Research Team

2026 AI chip market enters a "winner takes all" phase. NVIDIA holds 90%+ market share, AMD struggles at 10%, and Google/AWS/Huawei/Cerebras each occupy niche segments. But a group of AI chip startups are fighting to survive in the cracks — this article analyzes the 2026 status and future of Tenstorrent, SambaNova, Graphcore, Cambricon, Moore Threads, Biren, and Iluvatar.

2026 AI Chip Market Landscape

Top Tier: Duopoly + 5 Challengers

VendorMarket Share2025 RevenueStatus
NVIDIA90%+~$130BDominant
AMD5%~$5B (MI business)2nd place
Huawei1% (China 60%)~$3B (Ascend)China dominant
Google TPU1% (internal)N/AInternal use
AWS Trainium<1% (internal)N/AInternal use
Cerebras<1%$510MIPO pending
Groq (NVIDIA)<1%N/AIntegrated into NVIDIA

Startups

VendorFoundedTotal Funding2025 RevenueStatus
Tenstorrent2016$700M+~$30M (est.)Fundraising
SambaNova2017$1.1B+~$80M (est.)Revenue growing
Graphcore2016$700M+N/AAcquired by SoftBank (2024)
Cambricon2016A-share listed~$80MA-share $25B market cap
Moore Threads2020$500M+~$30MPreparing for IPO
Biren2019$700M+~$20MPreparing for IPO
Iluvatar2018$400M+~$15MListed on HKEX
Lightmatter2017$300M+~$5MSilicon photonics
Esperanto2014$120M<$5MRISC-V AI
Mythic2012$200M+<$5MEdge AI

Ten Startup Deep Dives

1. Tenstorrent: RISC-V Dark Horse

Tenstorrent is led by legendary chip designer Jim Keller (formerly designed AMD Zen / Apple A14 / Tesla FSD):

ItemDetails
FounderJim Keller (legendary architect)
Founded2016
HQToronto / Austin / Silicon Valley
Funding$700M+ (2024-12 valuation $2B)
2025 Revenue~$30M (est.)
Core ProductsWormhole n150/n300, Blackhole
FabTSMC 12nm + GlobalFoundries
SoftwareFully open-source (TT-Metalium)
CustomersLG, BOSCH, AutoDesk, RIKEN (Japan)
2026 PlanGrendel next-gen + strategic partnerships (speculated with OpenAI/AMD)

Tenstorrent Key Products

ProductProcessComputeMemoryPriceStatus
Wormhole n15012nm80 TOPS (FP8)12GB~$2K2023
Wormhole n30012nm160 TOPS (FP8)24GB~$4K2024
Blackhole p1506nm320 TOPS (FP8)16GB~$3K2025
Blackhole p3006nm800 TOPS (FP8)24GB~$6K2025
Grendel4nm (est.)1.5 POPS (FP8)32GBTBD2026-2027

Tenstorrent Key Advantages

  • Jim Keller's personal brand: Top-tier industry architect
  • Fully open-source software: TT-Metalium (GitHub 13K+ stars)
  • RISC-V ecosystem: Deep collaboration with SiFive / RISC-V International
  • Government / academic customers: RIKEN (Japan), multiple US universities

Tenstorrent Key Challenges

  • High TDP: Blackhole p300 ~150W (vs H100 700W but still weaker performance)
  • Weak ecosystem: PyTorch compatibility still improving
  • Low market awareness: Hard to acquire enterprise customers vs NVIDIA

2. SambaNova: Enterprise RDU Appliances

SambaNova is the representative of enterprise AI appliances:

ItemDetails
FoundersKunle Olukotun (Stanford professor) + 2 Stanford alumni
Founded2017
HQPalo Alto
Funding$1.1B+ (2021 valuation $5B)
2025 Revenue~$80M (est.)
Core ProductSN40L RDU (Reconfigurable Dataflow Unit)
FabTSMC 7nm
CustomersUS Government, Accenture, Hewlett Packard Enterprise
2026 PlanSN50 next-gen (larger RDU)

SambaNova SN40L Specifications

ItemSpec
ArchitectureRDU (Reconfigurable Dataflow Unit)
ProcessTSMC 7nm
RDU Cores1,040 tiles
HBM Capacity128 GB HBM3
HBM Bandwidth3.2 TB/s
FP16 Compute600 TFLOPS
BF16 Compute300 TFLOPS
TDP~600 W
Price~$150K / system

SambaNova Business Model

  • Does not sell chipssells appliances (SambaSystems)
  • SambaFlow software stack (Apache 2.0)
  • Focus on enterprise private deployment: Government, banking, telecom
  • 2025 Customers: US Air Force, Accenture, HPE

SambaNova Key Challenges

  • Small market share: Huge gap vs NVIDIA giant
  • Appliance model hard to scale: Each customer requires customization
  • 2024 20% layoffs (restructuring)

3. Graphcore: Acquired by SoftBank

Graphcore was a UK AI chip pioneer, but was acquired by SoftBank in 2024:

ItemDetails
FoundersNigel Toon + Simon Knowles
Founded2016
HQBristol (UK)
Funding$700M+ (2020 valuation $2.8B)
2024 Revenue~$30M (est.)
Core ProductsBow GC200 IPU + Bow Pod
FabTSMC 7nm
2024-10 AcquisitionAcquired by SoftBank (undisclosed, est. $600M)
2026 StatusSoftBank subsidiary, focused on Japan market

Graphcore Key Events

TimeEvent
2018Bow IPU debut
2020Peak valuation $2.8B
2022Revenue far below expectations
2023Multiple layoffs
2024-10SoftBank acquisition
2025Pivoted to Japan market (SoftBank Japan + Saudi G42)
2026SoftBank internal use + Japan National AI Strategy

Graphcore Future

  • No longer pursuing independent IPO
  • Integrated by SoftBank into ARM ecosystem
  • Bow Pod 128 remains flagship
  • Possible 2027 exit (SoftBank divestiture)

4. Cambricon: A-share $25B

Cambricon is China's first AI chip stock:

ItemDetails
Founded2016
Co-foundersChen Yunji, Chen Tianshi (brothers, CAS ICT)
Listed2020-07 STAR Market (A-share 688256)
Market Cap~$25B (2026-05)
2025 Revenue~$80M
Core ProductsSiyuan 290 / 590 / Siyuan 690 (next-gen)
FabSMIC
CustomersGovernment, telecom, internet

Cambricon Siyuan 590

ItemSpec
Process7nm
BF16 Compute480 TFLOPS
INT8 Compute960 TOPS
HBM Capacity64 GB
HBM Bandwidth2.4 Tbps
TDP~300 W
Price~$5K (est.)

Cambricon Challenges

  • Weak software ecosystem: MLU programming model gap vs CUDA
  • Market share squeezed by Huawei: Ascend 910C leading in compute since 2025
  • Poor profitability: Still loss-making

5. Moore Threads: A-share Under Review

Moore Threads is China's #2 GPU vendor:

ItemDetails
Founded2020
FounderZhang Jianzhong (former NVIDIA Global VP, China GM)
Funding$500M+
2025 Revenue~$30M
Core ProductsMTT S4000 / S5000
FabSMIC 7nm
A-share StatusApplied for STAR Market IPO late 2025

Moore Threads MTT S5000

ItemSpec
Process7nm (SMIC)
FP16 Compute250 TFLOPS
INT8 Compute500 TOPS
Memory32GB GDDR6X
Memory Bandwidth1.6 Tbps
TDP~300 W
Price~$3K

Moore Threads Challenges

  • Weak ecosystem: MUSA vs CUDA gap
  • Market share: Far below Huawei
  • A-share IPO pending

6. Biren: HKEX Under Review

Biren is China's #3 GPU vendor:

ItemDetails
Founded2019
FounderZhang Wen (Harvard PhD, former NVIDIA)
Funding$700M+
2025 Revenue~$20M
Core ProductBR104
FabSMIC 7nm
HKEX StatusApplied for HKEX listing late 2025

Biren BR104

ItemSpec
Process7nm (SMIC)
FP16 Compute300 TFLOPS
INT8 Compute600 TOPS
Memory32GB GDDR6
Memory Bandwidth1.6 Tbps
TDP~300 W
Price~$3K

Biren Key Events

  • 2023 US export controls limited advanced process access
  • 2024 IPO postponed (performance missed targets)
  • Late 2025 re-applied for HKEX listing

7. Iluvatar: HKEX Listed

Iluvatar is listed on HKEX:

ItemDetails
Founded2018
Listed2023 HKEX
Core ProductsTianhai 100 / Bi-150
FabSMIC 7nm
Market Cap~$500M

8. Lightmatter: Silicon Photonics Computing

Lightmatter is a silicon photonics pioneer:

ItemDetails
Founded2017
Funding$300M+
Core ProductEnvise silicon photonics AI accelerator
ProcessTSMC 5nm + proprietary photonics chip
CustomersMajor datacenters

Lightmatter Envise

ItemSpec
ArchitectureSilicon photonics + electronic hybrid
Compute1 PFLOP (FP16)
Power50% less than traditional GPU
2026 StatusCommercial pilot

9. Esperanto: RISC-V AI

Esperanto is a RISC-V AI accelerator representative:

ItemDetails
Founded2014
Funding$120M+
Core ProductET-SoC-1 (1,000+ RISC-V cores)
FabTSMC 7nm
2026 StatusKey customers: HPC / recommendation systems

10. Mythic: Edge AI

Mythic is an edge AI analog computing company:

ItemDetails
Founded2012
Funding$200M+
Core ProductM1076 analog AI chip
ProcessTSMC 40nm
2026 StatusPivoting / edge market

Startup Classification and Future

Category 1: Likely to Succeed (5-10 year IPO or acquisition)

VendorPathKey Support
TenstorrentIndependent IPO or acquisitionJim Keller + RISC-V + open source
CambriconContinue A-share listingChina national AI strategy
Moore ThreadsA-share IPOZhang Jianzhong (NVIDIA China background)

Category 2: Likely to Survive (Niche Markets)

VendorNiche Market
SambaNovaUS Government + enterprise private deployment
LightmatterSilicon photonics + datacenter low-power
EsperantoHPC + recommendation systems
Biren / IluvatarChina domestic substitution government market

Category 3: Likely to Disappear

VendorRisk
GraphcoreSoftBank subsidiary, uncertain future
MythicAnalog computing surpassed by digital
Small startupsTriple pressure: funding + customers + ecosystem

Investment Logic

Why Can Startups Survive?

  1. Niche markets: Government / defense / academic / specialized industries (NVIDIA doesn't address)
  2. Differentiated architecture: RISC-V / silicon photonics / analog / dataflow (NVIDIA won't pursue)
  3. Localization: China / Europe / Japan / India (data sovereignty + politics)
  4. Key customer lock-in: Tenstorrent + LG, SambaNova + US Air Force

Why Do Startups Die?

  1. NVIDIA ecosystem crushing: CUDA + cuDNN + TensorRT unbeatable
  2. High capital consumption: 7nm tape-out $30M+, 5nm $80M+
  3. Software migration costs: 6-12 months to migrate from CUDA to non-NVIDIA platforms
  4. High customer concentration: Losing a key customer = death

Detailed Product Pages

Summary

AI chip startups face "winner takes all":

  1. Tenstorrent most likely to succeed (Jim Keller + RISC-V + open source)
  2. SambaNova successfully pivoted (enterprise appliances + US Government)
  3. Graphcore acquired by SoftBank (fate transferred to Japan)
  4. Cambricon / Moore Threads / Biren benefit from China domestic substitution
  5. Lightmatter / Esperanto survive in niche markets
  6. Mythic and smaller companies unlikely to survive

Over the next 5 years, the AI chip industry will experience "big fish eat small fish" — startups will either IPO, get acquired, or disappear.