Silicon — GPUs, TPUs, ASICs
Fable 5 · data curated through July 2026
Layer 1 of five in the Five Layer AI Cake. 18 companies, 22 events.
What this layer is
Raw compute. Who makes the metal that every layer above depends on.
Layer 1 of five. Nothing sits below it — this is the floor of the stack, and the floor eats first. Above it sits Cloud & Compute, which pays it.
18 of the atlas's 103 companies capture value here. 8 of them also capture value on another layer, which makes them vertical integrators; 10 compete on this layer alone.
Vertical integrators on this layer (8)
These companies span two or more layers, so they hold margin at every floor they occupy.
| Company | Founded | Also on | Influence | Why it is here |
|---|---|---|---|---|
| Google Cloud | 2008 | Cloud & Compute · Foundation Models · Applications | 9/10 | TPUs + Gemini + Workspace. Only full-stack vertical. |
| Groq | 2016 | Cloud & Compute | 6/10 | LPU — record-setting inference latency. |
| Huawei (HiSilicon) | 1991 | Cloud & Compute | 6/10 | Ascend 910C/D — China's domestic GPU answer to NVIDIA under export controls. |
| Qualcomm | 1985 | Tooling & Orchestration | 6/10 | Bought Modular to fight CUDA with software. Silicon without a dev stack does not sell. |
| Tesla (Optimus, FSD) | 2003 | Applications | 6/10 | Dojo chips + Optimus + FSD — an AI company hiding in a carmaker. |
| Cerebras | 2016 | Cloud & Compute | 5/10 | Wafer-scale engine. Training alt to NVIDIA. |
| Intel | 1968 | Cloud & Compute | 5/10 | Gaudi chips, foundry ambitions; fell behind in AI. |
| SambaNova | 2017 | Cloud & Compute | 3/10 | Reconfigurable dataflow architecture. |
Pure plays on this layer (10)
| Company | Founded | Also on | Influence | Why it is here |
|---|---|---|---|---|
| NVIDIA | 1993 | — | 10/10 | GPU monopoly on AI training. CUDA moat. |
| TSMC | 1987 | — | 9/10 | Fabs every leading-edge AI chip on Earth. |
| AMD | 1969 | — | 7/10 | MI300X, ROCm — the credible challenger. |
| Broadcom | 1991 | — | 7/10 | Networking silicon for every hyperscaler. Tomahawk switches, custom XPUs. |
| Arm | 1990 | — | 6/10 | Architecture under every phone + edge device. |
| Arista Networks | 2004 | — | 5/10 | Data-center Ethernet. The anti-InfiniBand bet for AI clusters. |
| SMIC | 2000 | — | 5/10 | China's foundry champion, shipping 7nm without EUV. |
| Corning | 1851 | — | 4/10 | Fiber optic cable for sea-cable and data-center builds. 170-year-old picks-and-shovels. |
| Lightmatter | 2017 | — | 4/10 | Optical interconnect — photonic compute fabrics. Cluster-scale bet. |
| Tenstorrent | 2016 | — | 4/10 | Jim Keller’s RISC-V AI chips. Open-source ISA bet. |
What happened here (22 events)
| Year | Event | Magnitude | What happened |
|---|---|---|---|
| 1968 | Intel founded | 2 | The company that will miss the AI wave it enabled. (Intel) |
| 1987 | TSMC founded | 3 | The silent kingmaker of every AI chip era. (TSMC) |
| 1993 | NVIDIA founded | 3 | Gaming GPU company accidentally builds the AI stack. (NVIDIA) |
| 2006 | CUDA released | 3 | General-purpose GPU compute — the sleeper hit. (NVIDIA) |
| 2016 | Google TPU v1 | 2 | First mainstream AI ASIC. (Google Cloud) |
| 2020 | NVIDIA acquires Mellanox ($6.9B) | 2 | Networking = new moat. InfiniBand fabrics the AI era. (NVIDIA) |
| 2023 | NVIDIA H100 ramps | 3 | $30K a chip. Everyone is allocation-starved. (NVIDIA) |
| 2024 | NVIDIA Blackwell (B200) | 2 | The H100 successor. Trillion-param models in reach. (NVIDIA) |
| 2024 | NVIDIA becomes $3T company | 3 | Briefly the world’s most valuable company. Picks-and-shovels wins again. (NVIDIA) |
| 2025 | Groq + Cerebras land enterprise | 1 | Inference-cost war starts. NVIDIA still training king. (Groq) |
| 2025 | Huawei Ascend 910C ships at scale | 2 | US export controls backfire — China builds a domestic GPU supply chain in 3 years. (Huawei (HiSilicon)) |
| 2025 | Broadcom Tomahawk 6 / Arista 7800 AI Spine ship | 1 | Ethernet-first AI fabric. NVIDIA Spectrum-X gets real competition. (Broadcom) |
| 2026 | Blackwell Ultra + Rubin previews | 2 | Annual cadence locked in. Memory bandwidth is the new moat. (NVIDIA) |
| 2026 | Vera Rubin enters full production | 2 | Blackwell's successor in full production; RTX Spark pitches agents-at-home. (NVIDIA) |
| 2026 | Qualcomm buys Modular to fight CUDA | 2 | $3.9B for Mojo + MAX — buying the hardware-agnostic layer because silicon without a dev stack does not sell. (Qualcomm) |
| 2026 | NVIDIA books an $81.6B quarter | 1 | Revenue up 85%, data center up 92% — the boom's most expensive routine earnings beat. (NVIDIA) |
| 2026 | Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs | 1 | While Large Language Models (LLMs) rely on highly non-linear components, in this work we demonstrate that they exhibit fundamental linear... |
| 2026 | SLCA-GRPO: Resolving Cross-Segment Credit Misattribution in Tool-Calling RL | 1 | Tool-calling agents produce heterogeneous outputs, interleaving structured tool invocations with user-facing natural language summaries. ... |
| 2026 | Softmax Reparameterization for Output-Head Quantization | 1 | Large vocabularies make output heads a substantial inference cost in small language models. We propose softmax reparameterization, a post... |
| 2026 | LightMIS: Ultra-Lightweight Medical Image Segmentation Without a Stage-Wise Decoder | 1 | We present LightMIS, a scalable family of ultra-lightweight convolutional networks for 2D binary medical image segmentation without a lea... |
| 2026 | When Do Model Internals Help? Exploring the Role of Representation Engineering in LLM Safety | 1 | Reliable AI safeguards require both control mechanisms that reduce unsafe behavior and monitoring mechanisms that detect safety risks dur... |
| 2026 | Nereus: Adaptive Parallelism for LLM Post-Training | 1 | Reinforcement learning (RL) post-training for large language models (LLMs) coordinates multiple models across generation, inference, and ... |
Claims that rest on this layer
- Layers are value capture, not operations — Companies sit where they capture value, not where they operate. That single rule is what makes the cake a map instead of a directory.
The rest of the stack
- 1 · Silicon — you are here
- 2 · Cloud & Compute — Hyperscalers, neoclouds
- 3 · Foundation Models — The frontier labs
- 4 · Tooling & Orchestration — SDKs, agents, frameworks
- 5 · Applications — End-user products
Or read the full index, or open the interactive view filtered to this layer.
The Five Layer AI Cake is published by ConeApex, the trading name of Nirukta Digital XR Technologies Private Limited (CIN U62090TS2025PTC198275), Hyderabad, Telangana, India. The dataset is licensed CC BY 4.0 — cite it freely. See llms.txt for how to attribute it.