d-Matrix Corsair entered full production on June 9, 2026, manufactured by TSMC on its N6 (6nm) process in partnership with Alchip Technologies, with volume shipments beginning that summer to priority hyperscalers, neoclouds, and frontier AI labs (d-Matrix). That's a meaningfully different milestone than most "challenger chip" announcements this blog tracks. This one shipped, and d-Matrix published TFLOPS, memory bandwidth, and silicon area numbers to go with it, not just a roadmap slide.
What d-Matrix didn't do is claim Corsair replaces an NVIDIA H100 or B200. The company pitches it explicitly as a GPU co-processor, paired with NVIDIA Blackwell in a hybrid rack, handling the decode phase of inference while the GPU handles prefill (AI Weekly). That framing matters more than the spec sheet. This post covers what Corsair actually is, how its published numbers stack up against NVIDIA's shipping lineup, and, more practically, whether there's anywhere to rent it yet. Spoiler: not through a general-purpose cloud, not today.
What d-Matrix Corsair Actually Is: Digital In-Memory Compute Explained
Corsair is an inference-only accelerator built around Digital In-Memory Compute (DIMC), an architecture that does computation where the data already sits instead of shuttling model weights back and forth between memory and a separate processor (ServeTheHome). d-Matrix's own framing of the problem is blunt: "the biggest slowdown in inference isn't computation, it's moving data around" (Crypto Briefing). That's the same memory-bound bottleneck we cover in our AI memory wall guide: once a model is loaded, the constraint on decode-phase latency is usually how fast you can pull weights and KV cache off memory, not how many FLOPS the chip can theoretically push.
DIMC attacks that directly by embedding compute logic into small memory blocks, which ServeTheHome's Hot Chips 2025 coverage describes as 6MB "stashes" feeding the compute cores, paired with LPDDR5X memory placed extremely close to the processing elements. On low-latency batch inference, d-Matrix says that design gets time-per-output-token down to 2ms even for a model as large as Llama3-70B (ServeTheHome).
DIMC Architecture and the 6nm TSMC Silicon
Each Corsair PCIe card carries 6,400 mm² of silicon, with 4GB of integrated "Performance Memory" running at 300 TB/s of bandwidth and up to 512GB of off-chip "Capacity Memory" per card (Crypto Briefing). That two-tier memory split is the practical expression of DIMC: a small, extremely fast pool sitting right against compute for the hot working set, backed by a much larger pool for holding full model weights and KV cache. TSMC confirmed the manufacturing partnership directly: "TSMC is pleased to support d-Matrix's production ramp of the Corsair inference platform on our N6 process technology," said Lucas Tsai, VP at TSMC (PR Newswire).
Corsair ships in rack, server, and air-cooled PCIe card configurations. None of them need liquid cooling, which is a real departure from NVIDIA's Blackwell rack systems and simplifies deployment for data centers that haven't built out liquid cooling infrastructure (d-Matrix).
Corsair Isn't Positioned as a Standalone NVIDIA Replacement
This is the part easy to miss if you skim the launch coverage: d-Matrix isn't asking hyperscalers to rip out GPUs. Corsair sits in a rack next to NVIDIA Blackwell rather than trying to displace it, with the GPU handling the compute-heavy prefill phase of inference and Corsair taking over decode (AI Weekly). d-Matrix CEO Sid Sheth put it plainly at launch: "Corsair takes off from where the GPU leaves off, and this summer our customers will be able to experience the turbocharge d-Matrix brings at full rack scale" (PR Newswire).
That's a narrower, more defensible pitch than most "GPU killer" announcements make. It gives hyperscalers a second stack to run alongside Blackwell instead of an all-or-nothing swap, which is a real negotiating lever even if Corsair never runs a model end to end on its own.
AI Inference Chip Landscape: Where Corsair Fits Among Non-GPU Accelerators
Corsair is one entry in a crowded field of inference-only silicon making the same underlying bet: that a chip built specifically for serving, rather than a general-purpose GPU built to also handle training, can win on cost or latency for a slice of the inference market. Every one of them takes a different architectural path to get there.
- Qualcomm's AI200 bets on sheer memory capacity: 768GB of commodity LPDDR5X per card, roughly 10x an H100's HBM3, with no published FLOPS figure yet. Our Qualcomm AI200 vs NVIDIA breakdown covers that tradeoff in depth.
- Cerebras took the opposite memory strategy from Corsair, packing everything onto a single wafer-scale die with on-die SRAM instead of adding cheap off-chip capacity. See our Cerebras WSE-3 vs NVIDIA H100 comparison for how that plays out at different batch sizes.
- Etched's Sohu chip hard-codes transformer attention into silicon, giving up flexibility (no MoE, no multimodal support) for raw throughput on a fixed model shape. Our Etched Sohu vs NVIDIA breakdown walks through that tradeoff.
- NVIDIA's own Groq 3 LPU pursues an SRAM-first, deterministic-latency design distributed through a GPU-cloud-adjacent model. Our Groq 3 LPU explainer covers how that compares.
- Huawei's Ascend 950 is shaped by a different pressure entirely, export controls, not a pure architecture bet. Our Huawei Ascend 950 vs NVIDIA B300/B200 comparison covers that angle.
Corsair's differentiator inside that field is that it isn't pitched as a full replacement for any of the above. It's the only one of these explicitly designed to sit inside a GPU rack as a co-processor rather than to compete for the whole inference job.
Corsair vs H100/H200/B200: Claimed Throughput and Efficiency Numbers
d-Matrix's headline claim: a Corsair-plus-Blackwell configuration runs inference roughly 10x faster, 3x cheaper, and up to 5x more energy efficient than a standalone GPU (AI Weekly). That's a system-level claim about the pairing, not Corsair alone against an H100 head to head, and it's worth reading it that way rather than as a chip-for-chip swap number.
There's one independent data point behind it. Gimlet Labs tested a speculative-decoding workload and measured response time dropping from about 24 seconds to under 2 seconds when Corsair was paired with GPUs versus a GPU-only setup (PR Newswire). That's a real, named third party running a real workload, which puts Corsair a step ahead of most inference-chip launches at this stage. It's still one benchmark on one workload type, not a broad suite.
At Corsair's original 2024 unveiling, d-Matrix claimed up to 10x faster interactive speed, 3x better performance-per-TCO, and 3x greater energy efficiency versus GPU alternatives, plus 60,000 tokens/sec at 1ms/token for Llama3 8B on a single server (d-Matrix). The 2026 production numbers roughly track that original pitch, with the energy efficiency claim now pushed up to 5x in the paired configuration.
Spec Comparison Table
| Spec | d-Matrix Corsair (dual-card) | NVIDIA H100 SXM5 | NVIDIA H200 SXM5 | NVIDIA B200 SXM6 |
|---|---|---|---|---|
| Compute | 4,800 TFLOPs (MXINT8), 19,200 TFLOPs (MXINT4) | ~3,958 TFLOPS FP8 (sparse) | ~3,958 TFLOPS FP8 (sparse) | 9,000 TFLOPS FP4 (dense) |
| Performance memory | 4 GB at 300 TB/s (per card) | 80 GB HBM3 at 3.35 TB/s | 141 GB HBM3e at 4.8 TB/s | 192 GB HBM3e at 8 TB/s |
| Capacity memory | Up to 512 GB per card (off-chip) | Included in HBM3 above | Included in HBM3e above | Included in HBM3e above |
| Silicon | 6,400 mm² per card, TSMC N6 (6nm) | 80B transistors, HBM3 | 80B transistors, HBM3e | 208B transistors, HBM3e |
| Workload target | Inference only, paired with GPU | Training + inference | Training + inference | Training + inference |
| Cooling | Air-cooled (rack, server, PCIe) | Air or liquid (SXM) | Air or liquid (SXM) | Liquid |
| Positioning | Decode-phase co-processor alongside Blackwell | Standalone GPU | Standalone GPU | Standalone GPU |
| Cloud availability | Early access via Gimlet Cloud, 2H 2026 | Available now | Available now | Available now |
| Spheron pricing | Not applicable | From $5.07/hr on-demand (SXM5) | From $4.88/hr on-demand (SXM5) | From $5.37/hr spot (no on-demand SXM6 listed right now) |
Pricing fluctuates based on GPU availability. The prices above are based on 15 Aug 2026 and may have changed. Check current GPU pricing → for live rates.
What d-Matrix Hasn't Published: Independent Benchmarks and Full FLOPS Detail
Look at the FLOPS numbers in that table again: d-Matrix's are MXINT8 and MXINT4, a lower-precision integer format tuned for inference throughput, not the FP8 and FP4 floating-point formats NVIDIA publishes. The two aren't directly comparable without normalizing for precision and accuracy loss, and d-Matrix hasn't published a side-by-side accuracy comparison at matched precision. NVIDIA's numbers, by contrast, are independently benchmarked and have been in production workloads for multiple GPU generations.
The Gimlet Labs 24-second-to-2-second result is the only independently run test in the public record so far. Everything else, including the 10x/3x/5x pairing claim, comes from d-Matrix itself. That's typical for a chip a few months into full production, and it's a step ahead of most first-generation accelerator launches, which ship with no third-party numbers at all. But it means every comparison here, including the table above, sets a shipping, independently verified NVIDIA lineup against a newer platform with mostly vendor-reported figures. Treat the gap accordingly and watch for broader third-party benchmarks as more Corsair units reach customer hands.
Can You Rent Corsair Capacity Yet, or Is NVIDIA Still the Practical Choice
No general-purpose cloud provider lists Corsair as a self-serve rentable instance today. The only named consumption path is early access through Gimlet Cloud, targeted for the second half of 2026 (d-Matrix). If your team needs inference capacity this quarter, that's NVIDIA or AMD, full stop, and it will stay that way until a provider actually turns on Corsair instances.
d-Matrix AI Chip Pricing: What "Tens of Thousands of Dollars" Per Card Means for Cloud Rates
Corsair cards are priced in the tens of thousands of dollars per unit, according to reporting, though d-Matrix hasn't published an official list price or a cloud hourly rate (Crypto Briefing). That per-card figure alone doesn't tell you what a rented hour would cost. A cloud rate depends on utilization, power draw, amortization schedule, and whatever margin a provider builds in, none of which d-Matrix has disclosed. Until a provider posts a real hourly number, d-Matrix's 3x-cheaper claim is a projection against its own internal TCO model, not a price you can compare against an H100 on-demand rate today.
For the cost-per-token framework you'd need to actually test that 3x claim once real Corsair pricing exists, our AI inference cost economics playbook covers how to model serving cost independent of the sticker price on any given card.
Gimlet Cloud: The Only Named Path to Rentable Corsair Access
d-Matrix and Gimlet Labs announced their partnership on March 12, 2026, to integrate Corsair into Gimlet Cloud alongside GPUs. Availability to select customers is targeted for the second half of 2026, and early access can be requested now (d-Matrix). That's a single, named, waitlist-gated path, not a general-availability listing. If Corsair capacity matters to your 2026 planning, that request form is the only concrete lever to pull today; everything else is watching for a second provider to announce.
d-Matrix Corsair vs NVIDIA: Decision Framework
| Your situation | What to do |
|---|---|
| Need inference capacity to deploy this quarter | NVIDIA or AMD GPU cloud. Nothing Corsair-based is generally available yet. |
| Running agentic, coding, or voice workloads sensitive to decode-phase latency | Worth requesting Gimlet Cloud early access to track, but don't plan production capacity around a 2H 2026 target date. |
| Training or fine-tuning anywhere in the pipeline | NVIDIA. Corsair is inference-only; there's no path around that. |
| Evaluating whether a hybrid GPU-plus-accelerator rack fits a future deployment | Watch for independent benchmarks beyond the single Gimlet Labs result before committing infrastructure budget. |
| Building a 2026-2027 vendor diversification plan, not an immediate deployment | Track Corsair's Gimlet Cloud rollout and any second provider announcement; a waitlist isn't a rentable SKU. |
| Need portability across providers with no single-vendor lock-in | CUDA GPU cloud today. Corsair's software maturity outside d-Matrix's own stack is unproven at scale. |
The structural read: Corsair's pitch is credible precisely because it's narrow. d-Matrix isn't claiming to unseat NVIDIA, it's claiming to make an NVIDIA rack faster and cheaper on the decode phase specifically. That's a real product for hyperscalers who already run Blackwell at scale. It isn't yet a product a team renting GPU capacity can buy.
What This Means for Teams Renting GPUs Today
Nothing changes about your GPU cloud decision this year. Corsair isn't available through any general-purpose rental provider, and the one named path, Gimlet Cloud, is a waitlist targeting the second half of 2026, not a live SKU. If you're running inference workloads now, whether that's agentic pipelines, RAG, or standard LLM serving, H100 SXM5, H200 SXM5, and B200 SXM6 are available today, run your existing vLLM or TensorRT-LLM stack unchanged, and bill per minute with no reserved commitment. Check the Spheron API docs for provisioning details if you're sizing a serving cluster this quarter.
What's worth doing now is watching, not waiting. For the full GPU decision framework independent of Corsair's rollout, see our best GPU for AI inference guide. d-Matrix's DIMC bet is genuinely interesting engineering with one real third-party result behind it already, ahead of where most inference-chip challengers are at launch. Revisit it once Gimlet Cloud, or another provider, actually lists Corsair capacity you can rent by the hour.
If decode-phase latency is the bottleneck driving you to read about Corsair, H100, H200, and B200 instances are live on Spheron right now, no wait for a 2H 2026 early-access window required.
Rent H100 GPU → | H200 on Spheron → | View all GPU pricing →
Frequently Asked Questions
Corsair is d-Matrix's inference-only accelerator built on Digital In-Memory Compute architecture. It entered full production on June 9, 2026, manufactured by TSMC on the N6 (6nm) process with Alchip Technologies, and volume shipments to priority hyperscalers, neoclouds, and AI labs began that summer.
No. d-Matrix explicitly pitches Corsair as a co-processor that pairs with NVIDIA Blackwell GPUs in a hybrid rack, not a standalone replacement. The GPU handles the compute-heavy prefill phase of inference, and Corsair takes over the decode phase, where d-Matrix says the combination runs roughly 10x faster, 3x cheaper, and up to 5x more energy efficient than a GPU-only setup.
Not yet through any general-purpose cloud provider. The only named consumption path is early access through Gimlet Cloud, which d-Matrix and Gimlet Labs announced in March 2026, targeting availability to select customers in the second half of 2026. Teams that need inference capacity now are still choosing between NVIDIA and AMD GPU clouds.
d-Matrix hasn't published an official list price or a cloud hourly rate. Reporting puts individual Corsair cards in the tens of thousands of dollars per unit, in line with high-end data center accelerators, but there's no public per-hour rental figure to compare against NVIDIA's on-demand pricing yet.
d-Matrix's production dual-card Corsair platform claims 4,800 TFLOPs at MXINT8 precision and 19,200 TFLOPs at MXINT4 precision. Each PCIe card carries 6,400 mm² of silicon, 4GB of integrated Performance Memory at 300 TB/s bandwidth, and up to 512GB of off-chip Capacity Memory. An H100 SXM5, by comparison, has 80GB of HBM3 at 3.35 TB/s and about 3,958 TFLOPS FP8 with sparsity, a shipping and independently benchmarked baseline Corsair's own claims haven't been tested against yet.
