AI Compute Card Future Roadmap (2025–2028)
Based on vendor announcements, supply chain reports, and industry analysis. Actual release dates subject to official vendor confirmation. Last updated: 2026-07-25.
Status indicators: ✅ Launched / In Production ｜ 🔄 Upcoming ｜ 🔮 Outlook / Speculative ｜ ❌ Cancelled
2025 Q2–Q4 (Launched / In Production)
| Date | Product | Vendor | Key Details |
|---|---|---|---|
| 2025-04 | ✅ NVIDIA B200 | NVIDIA | Blackwell, 192GB HBM3e, 9 PFLOPS FP8 |
| 2025-04 | ✅ NVIDIA GB200 | NVIDIA | 2× B200 + Grace CPU, flagship LLM inference |
| 2025-05 | ✅ AMD MI355X | AMD | CDNA 4, 288GB HBM3e, 5 PFLOPS FP8 dense |
| 2025-06 | ✅ Google TPU v6e (Trillium) | 918 TFLOPS BF16, 32GB HBM, GA | |
| 2025-08 | ✅ Huawei Ascend 910C | Huawei | Dual-die 910B, 780 TFLOPS BF16, domestic production |
| 2025-09 | ✅ NVIDIA B300 Ultra | NVIDIA | 288GB HBM3e, 14 PFLOPS FP4, 1,400W liquid-cooled |
| 2025-12 | ✅ AWS Trainium 3 | AWS | 3nm, 5.7 PFLOPS FP8 dense, re:Invent GA |
| 2025-12 | ✅ Google TPU v7 (Ironwood) | 192GB HBM, 2,307 TFLOPS BF16, inference-optimized |
2026 H1 (Launched / In Production)
| Date | Product | Vendor | Key Details |
|---|---|---|---|
| 2026-Q1 | ✅ Huawei Ascend 950PR / 950DT | Huawei | 1 PFLOPS FP8, in-house HiBL/HiZQ HBM, in production |
| 2026-Q1 | ✅ Cambricon MLU690 | Cambricon | 2 PFLOPS FP8 dense, 192GB HBM3E, shipping |
| 2026-Q1 | ✅ AMD MI350X | AMD | CDNA 4, 192GB HBM3e, 2.5 PFLOPS FP16, available |
| 2026-02 | ✅ Moore Threads MTT S5000 | Moore Threads | 1,000 TFLOPS, 80GB GDDR6X, specs public |
| 2026-06 | ✅ NVIDIA Vera Rubin R200 | NVIDIA | Full production (announced at GTC Taipei) |
| 2026-06 | ✅ NVIDIA RTX Spark | NVIDIA | 20-core Grace + Blackwell GPU, 1 PFLOPS AI PC chip |
2026 H2 (Upcoming / Highly Anticipated)
| Date | Product | Vendor | Expected Specs |
|---|---|---|---|
| 2026-H2 | 🔄 NVIDIA Rubin R200 Shipping | NVIDIA | 288GB HBM4, 50 PFLOPS FP4, NVLink 6 |
| 2026-H2 | 🔄 NVIDIA Rubin NVL72 Rack | NVIDIA | 72 Rubin + 36 Vera, 1.8 EFLOPS FP4 |
| 2026-H2 | 🔄 AMD MI400 Series + Helios Rack | AMD | CDNA 5, TSMC 2nm, three SKUs (see below) |
| 2026-H2 | 🔄 AMD MI455X (Flagship) | AMD | 432GB HBM4, 40 PFLOPS FP4, Helios rack core |
| 2026-H2 | 🔄 AMD MI450X (Value) | AMD | 432GB HBM4, FP4/FP8 training + inference |
| 2026-H2 | 🔄 AMD MI430X (HPC) | AMD | 432GB HBM4, FP32/FP64 support, sovereign AI + HPC |
| 2026-H2 | 🔄 Qualcomm AI 200 | Qualcomm | Rack-scale AI inference, 768GB LPDDR/card, low TCO |
| 2026-Q4 | 🔄 Huawei Ascend 950DT Full Ramp | Huawei | Training mass production; WAIC 2026 showcased Atlas 950 SuperPoD (1024 chips, 1 EFLOPS FP8) |
| 2026-Fall | 🔄 NVIDIA RTX Spark Retail | NVIDIA | Asus / Dell / HP / Lenovo / Microsoft launch partners |
AMD MI400 SKU Matrix (Advancing AI 2026 / July 22 confirmed): Three models — MI455X (flagship training + inference) / MI450X (value training) / MI430X (HPC + FP64), all equipped with 432GB HBM4. Helios rack contains 72 MI455X + 18 EPYC Venice CPUs, TDP 225–245kW.
2027 (Outlook)
| Date | Product | Vendor | Expected Direction |
|---|---|---|---|
| 2027-H1 | 🔮 Qualcomm AI 250 | Qualcomm | Near-memory compute, 10× effective memory bandwidth, liquid-cooled rack |
| 2027 | 🔮 NVIDIA Rubin Ultra / Feynman | NVIDIA | Next-next-gen architecture, possibly wafer-scale |
| 2027 | 🔮 AMD MI500 series | AMD | CDNA 6, TSMC 2nm, HBM4E, co-developed with OpenAI |
| 2027 | 🔮 Google TPU v8 8t/8i | Training/inference split architecture, optical interconnect | |
| 2027 | 🔮 Cerebras WSE-4 (CS-4) | Cerebras | 3nm, ~5–6T transistors, ~200 PFLOPS BF16, post-IPO first gen |
| 2027-Q4 | 🔮 Huawei Ascend 960 | Huawei | FP8 ~2 PFLOPS, N+3 process |
| ❌ Intel Jaguar Shores | Intel | Cancelled (Xe-HPC + Gaudi converged architecture terminated) |
Intel Jaguar Shores: Originally planned for 2027–2028 as a Xe-HPC + Gaudi converged rack-scale system, now officially cancelled. Intel's AI GPU roadmap pivots to Crescent Island and subsequent products.
2028 (Long-term)
| Date | Product | Vendor | Expected Direction |
|---|---|---|---|
| 2028-Q4 | 🔮 Huawei Ascend 970 | Huawei | Third-gen Ascend, specs TBA |
Huawei Connect 2025 announced a three-generation roadmap: 950 series (2026) → 960 (2027-Q4) → 970 (2028-Q4).
Key Technology Trends
HBM4 / In-House HBM (2026–2027)
- HBM4 (NVIDIA Rubin / AMD MI400): 1,024-bit interface, 2.4+ TB/s/stack
- Huawei in-house HiBL/HiZQ: First Chinese enterprise to achieve in-house HBM mass production (950PR / 950DT)
- HBM4e: Late 2027, likely for Rubin Ultra
Personal AI Computing (2026 Emerging Trend)
- NVIDIA RTX Spark: AI PC chip, 1 PFLOPS, shipping Fall 2026
- Apple M5 Ultra: Local LLM inference, 192GB unified memory
- Qualcomm Snapdragon X Elite 2: 2027, AI PC continues to evolve
- Significance: AI compute extends from datacenters to personal devices
Datacenter AI Inference ASICs (2026 Emerging Trend)
- Qualcomm AI 200 (2026 commercial): Rack-scale AI inference, 768GB LPDDR/card, low TCO
- Qualcomm AI 250 (2027 commercial): Near-memory compute, 10× effective memory bandwidth
- Rack specs: Direct liquid cooling, 160kW rack-level power, Ethernet interconnect
- Significance: Traditional mobile chipmakers entering datacenter AI market, inference specialization trend
Wafer-Scale / Rack-Scale (2026+)
- NVIDIA Rubin NVL72: 72 GPU + 36 CPU, rack-scale AI factory
- AMD Helios: 432GB HBM4, 260 TB/s UALink interconnect
- Cerebras WSE-3: 4T transistor wafer-scale engine (current flagship)
- Huawei CloudMatrix 384: 384-chip interconnect, 750+ deployments; WAIC 2026 showcased Atlas 950 SuperPoD (1024 chips, 1 EFLOPS FP8)
Low-Precision Computing (FP4 / INT4)
- NVIDIA Rubin R200: 50 PFLOPS FP4 (sparse), 2.8× Blackwell
- AMD MI400: 40 PFLOPS FP4 (dense), Helios rack ~2 ExaFLOPS
- Huawei HiF8: Proprietary FP8 variant, precision approaching FP16
- Trend: FP8 → FP4 → INT4, trading precision for compute, models trained at ever-lower precision
Domestic Alternatives Accelerating (2026 Three-Pole Landscape)
| Vendor | Flagship | Compute | HBM | Ecosystem | Status |
|---|---|---|---|---|---|
| Huawei | Ascend 950DT | 1 PFLOPS FP8 | In-house HiZQ 144GB | CANN + MindSpore | ✅ In production |
| Cambricon | MLU690 | 2 PFLOPS FP8 | HBM3E 192GB | NeuWare + MindSpore | ✅ Shipping |
| Moore Threads | MTT S5000 | 1,000 TFLOPS | GDDR6X 80GB | MUSA + MUSIFY | ✅ Specs public |
Release Timeline (Visual)
timeline
title AI Compute Card Release Roadmap (2025-2028)
section 2025 H1
NVIDIA B200 : Launched
AMD MI355X : Launched
Google TPU v6e : GA
section 2025 H2
NVIDIA B300 Ultra : Launched
AWS Trainium 3 : GA
Google TPU Ironwood : Launched
Huawei Ascend 910C : In Production
section 2026 H1
NVIDIA Vera Rubin : Full Production
NVIDIA RTX Spark : Launched
Huawei Ascend 950PR/DT : In Production
Cambricon MLU690 : Shipping
section 2026 H2
Rubin NVL72 : Shipping
AMD MI400 Series : MI455X/MI450X/MI430X
AMD Helios Rack : Expected H2
Qualcomm AI 200 : Rack Inference
RTX Spark Retail : Fall
Ascend 950 SuperPoD : WAIC 2026
section 2027
Qualcomm AI 250 : Near-Memory
NVIDIA Feynman : Outlook
AMD MI500 : Outlook
Google TPU v8 : Outlook
Cerebras WSE-4 : Post-IPO
Huawei Ascend 960 : Roadmap
Intel Jaguar Shores : ❌ Cancelled
section 2028
Huawei Ascend 970 : Roadmap Q4
Data Sources & Updates
- Official vendor announcements (NVIDIA GTC Taipei 2026, Computex 2026, AMD Advancing AI 2026 (July 22), Huawei Connect 2025, WAIC 2026 (July 17), Qualcomm Oct 2025 announcement)
- Supply chain leaks (DigiTimes, Tom's Hardware, ServeTheHome)
- Industry analyst reports (TrendForce, Jon Peddie Research)
- Last updated: 2026-07-25, based on AMD Advancing AI 2026 (7/23 wrap-up: MI500 preview + OpenAI/Meta/Microsoft partnership details) / WAIC 2026 / GTC Taipei 2026 latest announcements
Found incorrect or missing information? Submit an Issue!
← Back to Home | Full Comparison Table → | TCO Calculator → | Industry News →