中文

Wang Kaixuan / 3D Vision & Robotics

Wang Kaixuan Blog

Personal blog on 3D vision, robotics, embodied AI and weekly notes.

Aug 17, 2026

AI Inference Chip Deep-Dive Research Report 2026

#投资调研 #AI芯片

The inference chip market is shifting from GPU dominance to diversified competition. This report analyzes inference workload characteristics, chip design trade-offs, and mainstream and emerging chip architectures, and lays out a cost-effectiveness-based investment framework.

Market

Exponential Growth in Inference Volume

Token consumption is the most direct underlying indicator of the inference market. The data below comes from public disclosures by top vendors, not third-party forecasts.

OpenAI API
15Btokens/min
Disclosed February 2026, ≈ 21.6P tokens/day
Google Cloud API
23Ttokens/day
April 2026, up 60% vs Q3 2025
China Market Total
140Ttokens/day
March 2026, a 1000× increase since early 2024

Global Inference Volume: From Hundreds of Billions to Hundreds of Trillions of tokens/day

The Current Inference Chip Landscape

The inference market is dominated by NVIDIA GPUs, but hyperscaler in-house chips are rapidly taking share. Below is an estimated distribution of global inference compute in 2026 (by share of token compute):

CampRepresentative ChipsEst. ShareChannelKey Advantage
NVIDIA GPU H100 / H200 / B200 ~74% Cloud / IDC / Enterprise Ecosystem, toolchain, interconnect
Hyperscaler in-house TPU / Trainium / MTIA ~18% Internal only / external GCP sales Best cost for internal traffic
AMD GPU MI300X / MI325X ~5% Cloud / some enterprises Large memory, low price
Dedicated inference chips Groq / Cerebras ~2% Dedicated API services Ultra-low latency / ultra-high throughput
Others Gaudi / Ascend ~1% Regional markets Domestic Chinese replacement / cost
Key Differences

The Fundamental Difference Between Inference and Training Workloads

DimensionTrainingInferenceImpact on Chips
Compute patternLarge-batch parallel (batch ≥ 256)Small batch or single sample (batch 1~16)Inference demands lower latency
Throughput vs latencyThroughput-first, second-level latency acceptableLatency-sensitive, P99 < 100msInference chips need low-latency scheduling
Memory accessBackprop requires storing intermediate activationsForward pass only, KV Cache grows linearlyThe inference bottleneck is memory bandwidth
PrecisionFP32 / BF16 / FP16INT8 / FP8 / INT4 acceptableInference can trade low precision for 2-4× throughput
Model fixednessWeights updated frequentlyWeights fixed, can be optimized offlineInference enables weight preloading and compilation optimization

Key Insight: Why Is the Inference Bottleneck Bandwidth?

During training, each token goes through forward + backward matrix multiplications twice — compute-intensive. Inference has two phases:

  • Prefill phase: processes all input prompt tokens at once — compute-intensive, with high compute utilization
  • Autoregressive (AR) phase: generates 1 token at a time, and must read the model weights + KV Cache (the K/V vectors of all past tokens) from HBM — bandwidth is the hard ceiling. For a 70B model + 8K context: KV Cache ≈ 21 GB, model weights ≈ 140 GB (FP16), ~161 GB of total memory reads; at 3 TB/s bandwidth, a single pass takes ~54 ms.

Multi-GPU Inference: Three Parallelism Strategies Compared

When a single GPU cannot meet performance or capacity requirements, multi-GPU parallelism is needed. The three mainstream strategies each have pros and cons for inference:

DimensionData ParallelTensor ParallelPipeline Parallel
Model distribution Each GPU stores a complete copy of the model Model split along intra-layer dimensions (Attention heads, MLP width) Model split vertically by layers (first N layers on GPU 1, last N on GPU 2)
Communication pattern ✅ No model communication ⚠️ All-reduce every layer
Communication ∝ hidden_dim × layers
⚠️ Activations passed between layers
Communication ∝ hidden_dim × (N-1)
Per-request latency = single-GPU latency
(no change)
= single-GPU latency / N + communication overhead
(can decrease, if communication is fast)
≈ single-GPU latency × (1 + bubble ratio)
(usually increases)
Total throughput Scales linearly, N×
(95%+ efficiency)
Depends on the communication bottleneck
Ideal < N×, in practice often 0.5-0.8N×
Depends on batch size
Approaches N× with large batches
Memory requirement Each GPU needs the full model
+ KV Cache
Each GPU needs only 1/N of the model
+ the full KV Cache
Each GPU needs only 1/N of the model
+ the full KV Cache
Best for • High-concurrency API serving
• Model fits on one GPU
• Throughput-first
• Model > single-GPU memory
• High-speed interconnect (NVLink)
• Low latency for a single user
• Training or large-batch inference
• Batch ≥ 8-16
• Throughput-first
Main limitation Model must be < single-GPU memory • Communication overhead ∝ layer count
• Needs high-speed interconnect (≥900GB/s)
Pipeline bubbles: low utilization at small batch
• Not suited to single-request inference
Inference suitability ⭐⭐⭐⭐⭐
Preferred choice
⭐⭐⭐⭐
Must-have for large models
⭐⭐
Large-batch scenarios only
💡 Real-world deployment lessons

1. Data parallel is the default for inference: 90% of API serving scenarios use data parallelism, because it has no communication overhead, is simple to implement, and scales throughput linearly.

2. Tensor parallel is for very large models: when the model > single-GPU memory (e.g. Llama 405B needs 4-8 GPUs), tensor parallelism is the only option. The key is inter-GPU link speed — the H100's NVLink at 900GB/s keeps communication overhead to 10-20ms, which is acceptable.

3. Pipeline parallel is rarely used for inference: pipeline bubbles waste 50-70% of compute at small batch sizes (1-4). Consider it only when batch ≥ 16 and the model is huge (> 100B). Google, Meta, and other big players internally combine tensor + pipeline parallelism (2D parallelism) to serve hundred-billion-parameter models.

4. Communication bandwidth is the lifeline of tensor parallelism: measured on Llama 70B tensor-parallel across 8 GPUs — with NVLink (900GB/s), communication overhead is ~8ms; with IB (200GB/s), it jumps to ~40ms and throughput drops 60%. This is why NVLink is standard inside DGX servers.
Mainstream Chips

The Mainstream Choices in Today's Inference Market

NVIDIA, AMD, Google, and Amazon hold 80-95% of the market. Below is a 2026 comparison of mainstream inference chips.

ChipFP16 ComputeHBM CapacityHBM BandwidthL2/L3 CacheBest Suited For
NVIDIA H100989 TFLOPS80 GB3.35 TB/s50 MBGeneral inference, mature ecosystem
NVIDIA H200989 TFLOPS141 GB4.8 TB/s50 MBSingle-GPU deployment of large models
NVIDIA B2002250 TFLOPS192 GB8 TB/s192 MBNext-generation flagship
AMD MI300X2600 TFLOPS192 GB5.3 TB/s256 MBMemory capacity advantage
Google TPU-8i~1200 TFLOPS192 GB7.4 TB/s~120 MBInference-dedicated
Groq LPU~750 TOPS (INT8)--230 MB SRAMUltra-low latency
Cerebras WSE-321.1 PFLOPS--44 GB SRAMWafer-scale chip
📦 On-chip SRAM vs HBM

Traditional GPU architectures (H100/MI300X):
• SRAM is small (50-256MB), used only as a temporary cache
• HBM is the main store; model weights and KV Cache both live in HBM
HBM bandwidth is the ceiling on inference performance

SRAM-centric architectures (Groq/Cerebras):
• Huge on-chip SRAM (230MB - 44GB); all model weights reside on-chip
No HBM bottleneck; memory access latency drops from 100ns to 1ns (100× speedup)
• Limitation: SRAM is expensive, so it only fits scenarios where the model < on-chip capacity
Emerging Players

Groq and Cerebras: Breakthroughs in Inference-Dedicated Chips

Groq and Cerebras use architectures fundamentally different from GPUs, designed specifically for low-latency inference.

Groq LPU (Language Processing Unit)

Core design idea: deterministic execution + extreme SRAM

Groq abandons the traditional GPU's thread scheduling and adopts TSP (Temporal Instruction Scheduling): the execution time of every operation is fixed at compile time, with zero scheduling overhead at runtime. The chip integrates 230MB of SRAM, keeping the KV Cache entirely on-chip.

Best for: small-batch / single-user inference, and latency-critical real-time applications (chatbots, code completion).

Limitations: SRAM capacity caps the maximum supported context (~32K); in April 2026 it was acquired by NVIDIA for $20 billion.

Cerebras WSE-3 (Wafer-Scale Engine)

Core design idea: wafer-scale chip + all-on-chip storage

Cerebras turns an entire wafer (300mm) into a single chip, integrating 44GB of SRAM and 900,000 cores. Model weights and the KV Cache all reside on-chip, eliminating the off-chip memory bottleneck.

Best for: models < 100B that need extreme throughput (API serving, batch generation). It IPO'd in May 2026, with its market cap briefly touching $80 billion.

Limitations: single-system cost >$2 million; energy efficiency far below GPUs.

Interactive Tool

Inference Performance Calculator

Pick a chip and a model to automatically compute the inference bottleneck and peak throughput.

Choose a Chip

Peak compute
-
HBM/SRAM
-
Memory bandwidth
-
On-chip cache
-

Model & Workload

70
8192
80
16K
1

Multi-GPU Cluster Configuration

1
300
KV Cache size
-
AR per-token latency
-
Theoretical peak throughput
-
Bottleneck analysis
-
📐 How the formulas work

KV Cache size = 2 (K+V) × context_len × hidden_dim × num_layers × 2 bytes (FP16)
AR latency = (model weights + KV Cache) / memory bandwidth
Peak throughput = (1000ms / latency) × batch_size
Compute density = (2 × model_params FLOPs) / (weights + KV Cache bytes)
Roofline analysis: if compute density < the chip's peak FLOPs/bandwidth ratio, the workload is bandwidth-bound

🔗 Multi-GPU parallel strategies

Data parallel: each GPU handles requests independently; throughput = single GPU × GPU count, with no model communication overhead. Best for high-concurrency API serving.
Tensor parallel: the model is split by layer, each GPU holds 1/N of the parameters and KV Cache, and every layer requires all-reduce communication. Best when a single model cannot fit on one GPU.
Pipeline parallel: the model is split into stages by layer, passing activations sequentially between them. Needs a large batch to fill the pipeline, otherwise bubbles are large.

Communication overhead (tensor-parallel all-reduce) = 4 × hidden_dim × layers bytes / inter-GPU bandwidth
NVLink bandwidth: H100 = 900GB/s (8 GPUs), IB/RoCE = 200-400GB/s

Investment View

Cost-Effectiveness Analysis and the Technology Roadmap Ahead

How Do You Evaluate Inference Cost-Effectiveness?

Inference cost-effectiveness = effective throughput / TCO. The key variables:

  • Effective throughput: not peak TPS, but sustained throughput under the target latency SLA (e.g. P99 < 200ms)
  • TCO: chip cost + power consumption (3-year depreciation) + interconnect / liquid cooling and other infrastructure + software maintenance costs
  • Utilization: resource utilization in multi-tenant scenarios. GPUs are general-purpose and can reach 70% utilization; dedicated chips often run below 40%

Technology Roadmap Outlook, 2026-2028

Short term (2026-2027): GPUs still dominate

NVIDIA H100/H200/B200 keep ruling on ecosystem strength, flexibility, and mature toolchains. AMD MI300X grabs share in cost-sensitive scenarios, but the ROCm ecosystem remains its bottleneck.

Investment take: NVIDIA is still the safest choice. AMD is the hedge.

Mid term (2027-2028): dedicated chips penetrate faster

Once inference workloads outweigh training, inference-dedicated chips (Groq, Cerebras, TPU-8i) climb from 5% to 15-20% penetration in specific scenarios.

Investment take: focus on players with deployments already at scale. Avoid ASICs that have not yet reached volume production.

Tech trend 1: the memory bandwidth arms race

The B200 offers 8TB/s of bandwidth; the next generation is expected to exceed 12TB/s. Bandwidth is inference performance — this is the core dimension of hardware iteration.

Tech trend 2: algorithm-hardware co-optimization

Algorithmic innovations like DeepSeek MLA, sparse attention, quantization (INT4/FP8), and KV Cache compression reduce dependence on bandwidth. Software-defined hardware efficiency becomes a differentiating competitive edge.

Investment Conclusion

Core judgments

1. NVIDIA's moat remains solid: its software ecosystem (CUDA, TensorRT), networking (NVLink), and toolchain maturity are barriers that cannot be closed within 3-5 years.

2. Dedicated chips are a complement, not a replacement: Groq/Cerebras have advantages in specific scenarios (low latency, fixed models), but cannot cover general-purpose inference needs. Their market share ceiling is ~15-20%.

3. Algorithmic innovation reduces hardware dependence: DeepSeek, Anthropic, and others use algorithmic innovation to make an H100 do work that used to require a B200. The bargaining power of software companies rises.

4. Watch bandwidth, not compute: the core metric of inference performance is shifting from TFLOPS to TB/s.

5. Hyperscaler in-house chips are splitting the mid-to-high-end market: Google TPU, AWS Trainium/Inferentia, and Azure Maia are all eroding NVIDIA's cloud inference market. In the long run, hyperscaler in-house chips could reach a 30-40% share.