02AHardware· as of 2026-10-15
AI accelerator roadmap
What ships next. 25 accelerators from 11 vendors that are announced but not in the spec table, each with the availability its source states. No throughput figures: those wait for a vendor document that gives dense rates.
Announced accelerators
Checked 2026-10-15. Ordered by the first month each stated window allows; sort by any column. Each window is the vendor’s own statement or, where marked, a press report.
Instinct MI440XAMD | 2026, with the MI400 series; AMD gave no separate date.For on-premises enterprise AI deployments. | Announced | ||
AI200Qualcomm | Commercially available in 2026, per Qualcomm.A rack-scale inference accelerator. | Announced | ||
Ascend 950PRHuawei | First quarter of 2026, on Huawei's September 2025 roadmap.First of Huawei's four-generation Ascend roadmap through 2028. | Announced | ||
MTIA 400Meta | Lab testing finished and heading to deployment, per Meta in March 2026.For Meta's own data centers. | Internal use | ||
Instinct MI350PAMD | Launched May 2026.A PCIe card: AMD's first PCIe Instinct accelerator since the MI210. | Launched | ||
TPU 8tGoogle | Later in 2026 for Google Cloud customers, as reported; Google has given no general-availability date.Eighth-generation TPU for training, scaling to 9,600 chips in one superpod. | Announced | ||
TPU 8iGoogle | Later in 2026 for Google Cloud customers, as reported; Google has given no general-availability date.Eighth-generation TPU for inference, connecting up to 1,152 chips. | Announced | ||
Groq 3 LPU (LPX rack)NVIDIA | Second half of 2026, as reported from GTC 2026; NVIDIA's May 2026 production release lists the LPX rack among its Vera Rubin racks.An inference accelerator from NVIDIA's December 2025 deal with Groq, with 500 MB of on-chip SRAM per LPU. | Announced | ||
Instinct MI430XAMD | Launched July 23, 2026, for HPC and sovereign-AI systems.An MI400-series part for scientific computing, slated for the Department of Energy's Discovery system at Oak Ridge, due in 2028. | Launched | ||
Crescent IslandIntel | Customer sampling in the second half of 2026, per Intel; no launch date.A data-center GPU. | Announced | ||
CS-4Cerebras | First shipments in the third quarter of 2026, per Cerebras.Built on three WSE-3 Turbo wafers. Cerebras has not announced a WSE-4; it targets its next wafer generation, in the CS-5, for 2027. | Announced | ||
OpenAI custom acceleratorOpenAI and Broadcom | Rack deployments from the second half of 2026 through the end of 2029, per OpenAI and Broadcom.10 GW of OpenAI-designed accelerators with Broadcom networking. | Announced | ||
Ascend 950DTHuawei | Fourth quarter of 2026, per Huawei, with the 8,192-chip Atlas 950 SuperPoD.Second chip of the 950 generation. | Announced | ||
Rubin CPXNVIDIA | End of 2026, per NVIDIA's September 2025 announcement.A GPU for long-context inference. It does not appear in NVIDIA's 2026 Vera Rubin platform pages or datasheet, and NVIDIA has not said whether the plan changed. | Announced | ||
Instinct MI500 seriesAMD | 2027, per AMD.Planned for AMD's Helios 500 rack. | Announced | ||
Trainium4AWS | AWS has given no date; press reports point to 2027.Previewed at re:Invent 2025 alongside Trainium3's general availability. | Announced | ||
Maia 300Microsoft | Production deployments from 2027, per Microsoft.Third-generation Azure Maia accelerator, designed for inference first. | Announced | ||
MTIA 450Meta | Mass deployment in early 2027, per Meta.For Meta's own data centers. | Internal use | ||
MTIA 500Meta | Mass deployment in 2027, per Meta.For Meta's own data centers. | Internal use | ||
AI250Qualcomm | Commercially available in 2027, per Qualcomm.A rack-scale inference accelerator; Qualcomm says its memory design gives more than 10 times the effective memory bandwidth. | Announced | ||
Rubin UltraNVIDIA | Second half of 2027, per NVIDIA at GTC 2025 and again at GTC 2026.Built into NVIDIA's Kyber rack. NVIDIA has published no datasheet. | Announced | ||
Ascend 960Huawei | Fourth quarter of 2027, per Huawei, with the 15,488-chip Atlas 960 SuperPoD.Third generation on Huawei's roadmap. | Announced | ||
FeynmanNVIDIA | 2028, on NVIDIA's GTC 2025 roadmap.NVIDIA's GTC 2026 news pairs the platform with a Rosa CPU and an LP40 inference chip. | Announced | ||
Ascend 970Huawei | Fourth quarter of 2028, per Huawei.Huawei said its specifications were still being finalized. | Announced | ||
Jaguar ShoresIntel | No date given by Intel.A rack-scale system Intel named when it made Falcon Shores an internal test chip. | Announced |
Already in the spec table
The hardware spec table holds 23 accelerators with dense rates read from vendor documents; its newest launches are Instinct MI455X (Oct 2026), Rubin GPU (Aug 2026), Maia 200 (Jan 2026). A chip moves there from this page once its vendor publishes the figures.
Background reading: AI Accelerator on AI Wiki. AI Wiki is run by the same team as AI Compute Tracker.