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
| Vendor | Market Share | 2025 Revenue | Status |
|---|
| NVIDIA | 90%+ | ~$130B | Dominant |
| AMD | 5% | ~$5B (MI business) | 2nd place |
| Huawei | 1% (China 60%) | ~$3B (Ascend) | China dominant |
| Google TPU | 1% (internal) | N/A | Internal use |
| AWS Trainium | <1% (internal) | N/A | Internal use |
| Cerebras | <1% | $510M | IPO pending |
| Groq (NVIDIA) | <1% | N/A | Integrated into NVIDIA |
Startups
| Vendor | Founded | Total Funding | 2025 Revenue | Status |
|---|
| Tenstorrent | 2016 | $700M+ | ~$30M (est.) | Fundraising |
| SambaNova | 2017 | $1.1B+ | ~$80M (est.) | Revenue growing |
| Graphcore | 2016 | $700M+ | N/A | Acquired by SoftBank (2024) |
| Cambricon | 2016 | A-share listed | ~$80M | A-share $25B market cap |
| Moore Threads | 2020 | $500M+ | ~$30M | Preparing for IPO |
| Biren | 2019 | $700M+ | ~$20M | Preparing for IPO |
| Iluvatar | 2018 | $400M+ | ~$15M | Listed on HKEX |
| Lightmatter | 2017 | $300M+ | ~$5M | Silicon photonics |
| Esperanto | 2014 | $120M | <$5M | RISC-V AI |
| Mythic | 2012 | $200M+ | <$5M | Edge 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):
| Item | Details |
|---|
| Founder | Jim Keller (legendary architect) |
| Founded | 2016 |
| HQ | Toronto / Austin / Silicon Valley |
| Funding | $700M+ (2024-12 valuation $2B) |
| 2025 Revenue | ~$30M (est.) |
| Core Products | Wormhole n150/n300, Blackhole |
| Fab | TSMC 12nm + GlobalFoundries |
| Software | Fully open-source (TT-Metalium) |
| Customers | LG, BOSCH, AutoDesk, RIKEN (Japan) |
| 2026 Plan | Grendel next-gen + strategic partnerships (speculated with OpenAI/AMD) |
Tenstorrent Key Products
| Product | Process | Compute | Memory | Price | Status |
|---|
| Wormhole n150 | 12nm | 80 TOPS (FP8) | 12GB | ~$2K | 2023 |
| Wormhole n300 | 12nm | 160 TOPS (FP8) | 24GB | ~$4K | 2024 |
| Blackhole p150 | 6nm | 320 TOPS (FP8) | 16GB | ~$3K | 2025 |
| Blackhole p300 | 6nm | 800 TOPS (FP8) | 24GB | ~$6K | 2025 |
| Grendel | 4nm (est.) | 1.5 POPS (FP8) | 32GB | TBD | 2026-2027 |
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:
| Item | Details |
|---|
| Founders | Kunle Olukotun (Stanford professor) + 2 Stanford alumni |
| Founded | 2017 |
| HQ | Palo Alto |
| Funding | $1.1B+ (2021 valuation $5B) |
| 2025 Revenue | ~$80M (est.) |
| Core Product | SN40L RDU (Reconfigurable Dataflow Unit) |
| Fab | TSMC 7nm |
| Customers | US Government, Accenture, Hewlett Packard Enterprise |
| 2026 Plan | SN50 next-gen (larger RDU) |
SambaNova SN40L Specifications
| Item | Spec |
|---|
| Architecture | RDU (Reconfigurable Dataflow Unit) |
| Process | TSMC 7nm |
| RDU Cores | 1,040 tiles |
| HBM Capacity | 128 GB HBM3 |
| HBM Bandwidth | 3.2 TB/s |
| FP16 Compute | 600 TFLOPS |
| BF16 Compute | 300 TFLOPS |
| TDP | ~600 W |
| Price | ~$150K / system |
SambaNova Business Model
- Does not sell chips — sells 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:
| Item | Details |
|---|
| Founders | Nigel Toon + Simon Knowles |
| Founded | 2016 |
| HQ | Bristol (UK) |
| Funding | $700M+ (2020 valuation $2.8B) |
| 2024 Revenue | ~$30M (est.) |
| Core Products | Bow GC200 IPU + Bow Pod |
| Fab | TSMC 7nm |
| 2024-10 Acquisition | Acquired by SoftBank (undisclosed, est. $600M) |
| 2026 Status | SoftBank subsidiary, focused on Japan market |
Graphcore Key Events
| Time | Event |
|---|
| 2018 | Bow IPU debut |
| 2020 | Peak valuation $2.8B |
| 2022 | Revenue far below expectations |
| 2023 | Multiple layoffs |
| 2024-10 | SoftBank acquisition |
| 2025 | Pivoted to Japan market (SoftBank Japan + Saudi G42) |
| 2026 | SoftBank 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:
| Item | Details |
|---|
| Founded | 2016 |
| Co-founders | Chen Yunji, Chen Tianshi (brothers, CAS ICT) |
| Listed | 2020-07 STAR Market (A-share 688256) |
| Market Cap | ~$25B (2026-05) |
| 2025 Revenue | ~$80M |
| Core Products | Siyuan 290 / 590 / Siyuan 690 (next-gen) |
| Fab | SMIC |
| Customers | Government, telecom, internet |
Cambricon Siyuan 590
| Item | Spec |
|---|
| Process | 7nm |
| BF16 Compute | 480 TFLOPS |
| INT8 Compute | 960 TOPS |
| HBM Capacity | 64 GB |
| HBM Bandwidth | 2.4 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:
| Item | Details |
|---|
| Founded | 2020 |
| Founder | Zhang Jianzhong (former NVIDIA Global VP, China GM) |
| Funding | $500M+ |
| 2025 Revenue | ~$30M |
| Core Products | MTT S4000 / S5000 |
| Fab | SMIC 7nm |
| A-share Status | Applied for STAR Market IPO late 2025 |
Moore Threads MTT S5000
| Item | Spec |
|---|
| Process | 7nm (SMIC) |
| FP16 Compute | 250 TFLOPS |
| INT8 Compute | 500 TOPS |
| Memory | 32GB GDDR6X |
| Memory Bandwidth | 1.6 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:
| Item | Details |
|---|
| Founded | 2019 |
| Founder | Zhang Wen (Harvard PhD, former NVIDIA) |
| Funding | $700M+ |
| 2025 Revenue | ~$20M |
| Core Product | BR104 |
| Fab | SMIC 7nm |
| HKEX Status | Applied for HKEX listing late 2025 |
Biren BR104
| Item | Spec |
|---|
| Process | 7nm (SMIC) |
| FP16 Compute | 300 TFLOPS |
| INT8 Compute | 600 TOPS |
| Memory | 32GB GDDR6 |
| Memory Bandwidth | 1.6 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:
| Item | Details |
|---|
| Founded | 2018 |
| Listed | 2023 HKEX |
| Core Products | Tianhai 100 / Bi-150 |
| Fab | SMIC 7nm |
| Market Cap | ~$500M |
8. Lightmatter: Silicon Photonics Computing
Lightmatter is a silicon photonics pioneer:
| Item | Details |
|---|
| Founded | 2017 |
| Funding | $300M+ |
| Core Product | Envise silicon photonics AI accelerator |
| Process | TSMC 5nm + proprietary photonics chip |
| Customers | Major datacenters |
Lightmatter Envise
| Item | Spec |
|---|
| Architecture | Silicon photonics + electronic hybrid |
| Compute | 1 PFLOP (FP16) |
| Power | 50% less than traditional GPU |
| 2026 Status | Commercial pilot |
9. Esperanto: RISC-V AI
Esperanto is a RISC-V AI accelerator representative:
| Item | Details |
|---|
| Founded | 2014 |
| Funding | $120M+ |
| Core Product | ET-SoC-1 (1,000+ RISC-V cores) |
| Fab | TSMC 7nm |
| 2026 Status | Key customers: HPC / recommendation systems |
10. Mythic: Edge AI
Mythic is an edge AI analog computing company:
| Item | Details |
|---|
| Founded | 2012 |
| Funding | $200M+ |
| Core Product | M1076 analog AI chip |
| Process | TSMC 40nm |
| 2026 Status | Pivoting / edge market |
Startup Classification and Future
Category 1: Likely to Succeed (5-10 year IPO or acquisition)
| Vendor | Path | Key Support |
|---|
| Tenstorrent | Independent IPO or acquisition | Jim Keller + RISC-V + open source |
| Cambricon | Continue A-share listing | China national AI strategy |
| Moore Threads | A-share IPO | Zhang Jianzhong (NVIDIA China background) |
Category 2: Likely to Survive (Niche Markets)
| Vendor | Niche Market |
|---|
| SambaNova | US Government + enterprise private deployment |
| Lightmatter | Silicon photonics + datacenter low-power |
| Esperanto | HPC + recommendation systems |
| Biren / Iluvatar | China domestic substitution government market |
Category 3: Likely to Disappear
| Vendor | Risk |
|---|
| Graphcore | SoftBank subsidiary, uncertain future |
| Mythic | Analog computing surpassed by digital |
| Small startups | Triple pressure: funding + customers + ecosystem |
Investment Logic
Why Can Startups Survive?
- Niche markets: Government / defense / academic / specialized industries (NVIDIA doesn't address)
- Differentiated architecture: RISC-V / silicon photonics / analog / dataflow (NVIDIA won't pursue)
- Localization: China / Europe / Japan / India (data sovereignty + politics)
- Key customer lock-in: Tenstorrent + LG, SambaNova + US Air Force
Why Do Startups Die?
- NVIDIA ecosystem crushing: CUDA + cuDNN + TensorRT unbeatable
- High capital consumption: 7nm tape-out $30M+, 5nm $80M+
- Software migration costs: 6-12 months to migrate from CUDA to non-NVIDIA platforms
- High customer concentration: Losing a key customer = death
Detailed Product Pages
Summary
AI chip startups face "winner takes all":
- Tenstorrent most likely to succeed (Jim Keller + RISC-V + open source)
- SambaNova successfully pivoted (enterprise appliances + US Government)
- Graphcore acquired by SoftBank (fate transferred to Japan)
- Cambricon / Moore Threads / Biren benefit from China domestic substitution
- Lightmatter / Esperanto survive in niche markets
- 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.