AI chip memory: which startup is ahead?

Which AI chip memory startups actually belong in this race?

The serious AI chip memory startup field contains eight companies, but they attack three different problems: moving AI computation closer to memory, expanding or pooling memory around accelerators, and replacing inefficient memory blocks inside chips. We include d-Matrix, Axelera AI, MemryX and EnCharge AI as memory-centric accelerator companies; Panmnesia and XCENA as CXL or computational-memory companies; and ZeroPoint Technologies and RAAAM Memory Technologies as memory-IP specialists.

We exclude conventional HBM makers such as SK hynix, Samsung and Micron because they are established semiconductor companies, not startups. We also leave Lightmatter and Ayar Labs out of the core ranking because their main product is optical chip-to-chip connectivity, even though it relieves the same data-movement bottleneck. Lightmatter has raised approximately $850 million and reached a $4.4 billion valuation, and Ayar Labs said in September 2026 that its funding for the year had reached $650 million. Both are important neighbors, not direct computational-memory competitors. Celestial AI is excluded because Marvell completed its acquisition of the company. Positron AI is a closer neighbor: it builds memory-first inference systems, raised $875 million in two tranches at a $5 billion valuation in September 2026 and says more than 50 of its Atlas racks are deployed at Oracle Cloud Infrastructure. It sells complete inference systems rather than memory technology, so we treat it as an adjacent rival.

The funding figures need caution. Axelera’s total includes equity, grants and venture debt, while Panmnesia and RAAAM combine private investment with public grants. ZeroPoint has announced a €5 million Series A but has not published a reliable cumulative total. d-Matrix’s $450 million is the company’s own total since its founding; Reuters has described the same sum as raised in 2025 alone.

Startup	What it is building	Main market	Cumulative disclosed funding
d-Matrix	Digital in-memory accelerators and rack-scale clusters for generative-AI inference	Data centers	$450M (company-stated); new round at a higher valuation reported, not announced
Axelera AI	Digital in-memory and dataflow accelerators (Metis, Europa)	Edge AI, now shipping into enterprise data centers	More than $450M in total capital
XCENA	CXL computational memory combining pooled memory with near-data processing	Data centers and HPC	$185M
EnCharge AI	Analog in-memory AI accelerators	PCs, workstations and edge systems	More than $144M
Panmnesia	CXL controllers, switches and GPU memory expansion systems	Data centers	More than $80M
MemryX	Low-power at-memory dataflow accelerators	Edge devices and edge servers	Approximately $63M
RAAAM Memory Technologies	GCRAM embedded-memory IP intended to replace SRAM	Semiconductor IP	More than $24M
ZeroPoint Technologies	Hardware compression IP for HBM, DDR and CXL memory	Semiconductor IP and data centers	Cumulative total undisclosed; €5M Series A announced

Is there already a clear AI chip memory startup leader?

Yes, d-Matrix is the clear leader, though not a runaway one. It is the only company in the field that combines a data-center generative-AI product in full production, multi-rack systems shipping to customers, a named commercial deployment and workload-level proof that its memory-centric design speeds up inference.

The 2026 milestones stack up. In August, d-Matrix bought Wallaroo.ai for deployment and orchestration software, its second acquisition in four months after GigaIO’s data-center business in April. In September, it joined NVIDIA’s NVLink Fusion program. In October, it announced Squadron, clusters of up to 1,024 Corsair XPUs that it says are now shipping to priority customers. Clusters already shipping to priority customers beat roadmaps, and we compare who is ahead, and who is falling behind, in our report on AI chips.

The closest challenger is Axelera AI. Axelera says it has more than 600 customers, has a shipped first-generation product, and began shipping its second chip, Europa, in September 2026. Its strongest traction is still at the edge, where deployments are smaller and the memory problem differs from large generative-AI models. d-Matrix is attacking the harder and potentially more valuable data-center decode bottleneck head-on.

d-Matrix and Axelera each report roughly $450 million of capital, while the next best-funded direct competitor, XCENA, has raised $185 million.

The rest of the field is split among narrower leaders: XCENA and Panmnesia in CXL memory, ZeroPoint in memory compression and RAAAM in embedded memory. MemryX remains the most mature smaller challenger.

Which AI memory startups have products customers can actually use?

d-Matrix, Axelera and MemryX have crossed the key product-maturity threshold. Each has production hardware, working software and a form factor customers can plug into systems they already run.

d-Matrix has gone furthest. Corsair is in full production, and in October 2026 the company made Corsair servers and eight-server SquadRacks generally available and said full Squadron clusters are shipping to priority customers. The clusters are air-cooled, so they fit existing data centers without liquid cooling. Axelera now sells two generations: Metis at the edge and Europa, a 629-TOPS chip at 45 watts by the company’s figures, shipping since September 2026 in validated Dell XE5 and Supermicro servers. MemryX sells the MX3 and its Cascade accelerators in M.2, PCIe and USB form factors.

The second group has working technology but no comparable production scale. XCENA unveiled an MX1 production lineup in August 2026 but has not disclosed volume shipments. Panmnesia is supplying pre-release CXL 3.2 switch silicon and says it plans mass production. EnCharge has launched the 200-TOPS EN100 but has not announced named volume deployments. ZeroPoint has licensed its IP to a global semiconductor manufacturer, but the end product and royalty scale remain undisclosed. RAAAM is still at the test-chip stage, now with GlobalFoundries.

Startup	Most advanced public stage	What is still missing
d-Matrix	Full production; Squadron clusters shipping to priority customers	Shipment volume, revenue and repeat-order data
Axelera AI	Deployments with more than 600 customers; Europa shipping in Dell and Supermicro servers	Customer mix, recurring revenue and deployment sizes
MemryX	Mass-production MX3 hardware; live edge deployments with Lenovo	Broad customer-count and revenue disclosure
ZeroPoint Technologies	IP licensed to a major semiconductor manufacturer	Named customer, product volumes and royalties
EnCharge AI	EN100 launched	Named production customers and shipment data
Panmnesia	Pre-release silicon, pilot systems and announced mass-production plan	Paid volume deployments
XCENA	MX1 production lineup; SK hynix product built on MX1 in validation	Volume shipments and public benchmarks
RAAAM Memory Technologies	Licensed IP, test chips and a joint GCRAM program with GlobalFoundries	Mass-produced customer silicon

Who has converted AI memory technology into the strongest commercial traction?

Axelera has the broadest disclosed customer base, but d-Matrix has the strongest evidence of commercially important data-center adoption. A raw customer count hides that difference.

Axelera says it has deployments with more than 600 customers across defense, robotics, drones, retail and security. It also says it has signed supply contracts worth tens of millions of dollars with European AI factories and has a sales pipeline above $1.5 billion. The contracts are a real step; the pipeline is a company estimate, not revenue. Axelera still does not say how many customers run large deployments or reorder.

d-Matrix discloses fewer customers, but the quality of its evidence is higher. Parasail runs Corsair alongside NVIDIA infrastructure in what the companies call one of the first commercial-scale heterogeneous inference deployments, Gimlet Labs has integrated Corsair into its cloud, and Squadron clusters are now going to priority customers. The NVLink Fusion deal gives d-Matrix a route into NVIDIA-based AI factories that no rival in this field has.

MemryX has a concrete commercial foothold: in September 2026 it signed a memorandum of understanding with Lenovo to expand edge AI in Saudi Arabia, building on joint deployments already running there. It also signed a partnership with Saudi integrator Ebttikar in July and said in August that Express LUCK had chosen its chips for factory-safety systems. ZeroPoint still relies on one license with an unnamed manufacturer. XCENA has SK hynix testing a product built on its chip, while EnCharge has not named a volume EN100 customer.

Is d-Matrix still ahead now that Squadron clusters are shipping?

Yes. d-Matrix is ahead in data-center AI inference memory because Corsair has moved from a card to a full rack-scale product, and the company is shipping it against the decode bottleneck that matters most in generative AI.

Corsair uses digital in-memory computation to keep model operations close to the data instead of moving weights through a conventional memory hierarchy. It started as a standard PCIe card. In October 2026 d-Matrix added the Squadron system: up to 16 racks and 1,024 Corsair XPUs, linked by its own JetStream network cards and built with Arista, Broadcom and Supermicro. GPUs keep the compute-heavy prefill, while Corsair handles latency-sensitive token generation.

That split lowers adoption risk because buyers keep their NVIDIA fleet. In September 2026 d-Matrix agreed a multi-year roadmap with NVIDIA to put its next chip, Raptor, into NVIDIA’s MGX rack design through NVLink Fusion, next to Vera Rubin GPUs. Raptor stacks compute directly on custom DRAM; d-Matrix presented it at Hot Chips in August 2026 with a claimed 100 TB/s of memory bandwidth per card. It is due to tape out by the end of 2026, and MGX rack availability is planned for late 2027.

The performance evidence keeps building. Gimlet Labs reported a two-to-five-times end-to-end speed-up at equal energy when running a speculative decoder on Corsair for gpt-oss-120b. At the OCP Global Summit in October 2026, d-Matrix split attention and expert layers between GPUs and Corsair racks and said the setup reached 845 tokens per second per user. Tech Times reported time per output token falling from 2.59 to 1.28 milliseconds versus a standalone H100 on GPT-OSS-20B.

The open question is still scale. d-Matrix has not said how many cards or clusters it has shipped, what customers pay or whether early deployments are expanding.

Can Axelera AI’s Europa chip take it beyond edge AI?

Possibly, but not yet. Axelera AI is clearly the strongest AI memory startup at the edge, and Europa gives it a credible product for enterprise servers. It does not yet give it a data-center position comparable to d-Matrix.

The edge lead rests on Metis. Axelera reports 214 TOPS from one Metis unit, about 15 TOPS per watt and 3,200 frames per second on ResNet-50, and an independent developer reproduced and then beat that throughput while measuring power separately. Europa doubles the processing cores to eight, adds 16 RISC-V cores for pre- and post-processing and is built on Samsung’s 5nm process. Axelera claims up to six times more tokens per second per watt than GPU-based systems, a figure based on its own testing against competitors’ published data.

Commercial breadth is still Axelera’s main advantage. More than 600 customers means repeated procurement, installation and software onboarding. Validated systems from Dell, Supermicro, HPE and Lenovo widen its sales channel. Its February 2026 round of more than $250 million took total capital through equity, grants and venture debt above $450 million, and its CEO told a Japanese outlet in October that an IPO in North America within three to five years is the goal. Axelera's round took its total capital above $450 million; we look at who's getting funded, and by whom, in our report covering AI chips.

Axelera stays second because Europa targets enterprise inference on smaller and mid-size models, not the frontier-scale decode work d-Matrix serves. The ranking could flip if its AI factory contracts grow into large repeat orders.

Which AI memory startup has proved the best performance?

No AI memory startup has proved universal performance leadership because public results cover different models, power envelopes and settings. d-Matrix has the strongest evidence for large-model decode, Axelera has the strongest reproducible edge-vision throughput, and MemryX has the best independent evidence for usability and low-power integration.

d-Matrix’s results are the most commercially relevant for generative AI. Gimlet’s two-to-five-times end-to-end gain at equal energy measures response completion, not theoretical arithmetic. The October 2026 OCP demo adds a second, company-run result on a mixture-of-experts model. Both apply to specific heterogeneous setups and say nothing about prefill, training or image generation.

Axelera’s evidence is narrower but more independently reproducible. Metis has been tested on a standard ResNet-50 workload with external power measurement. Europa’s 629 TOPS and its Llama comparisons are still company figures. MemryX received a hands-on evaluation from BDTI, which found the MX3 module delivered about 14 to 24 TFLOPS depending on configuration and was unusually easy to install and use.

EnCharge’s EN100 still rests on a headline 200 TOPS without a public customer workload comparison. With no shared benchmark across the field, performance leadership remains workload-specific.

Which AI memory startup offers customers the best economics?

d-Matrix has the strongest evidence of better data-center inference economics, while Axelera has the clearest edge and enterprise price-performance story. Neither publishes enough pricing, utilization or payback data to prove a universal cost advantage.

d-Matrix claims Corsair delivers about three times better cost-performance and three-to-five-times better energy efficiency than GPU-only inference. Gimlet’s results support the mechanism, and Squadron’s air-cooled design avoids the cost of liquid cooling or new buildings. Purchase prices and utilization are still not public.

Axelera publishes a claimed 16.4 ResNet-50 frames per second per dollar for Metis, which typically runs below 15 watts. For Europa it claims three to five times better performance per dollar on Llama models, using its own tests against rivals’ published data. Neither figure includes full system cost, software integration or support.

MemryX fits where power and simplicity matter more than peak throughput. One MX3 chip typically draws 0.6 to 2 watts, and a four-chip M.2 module provides up to 24 TFLOPS. Its ease of deployment also cuts engineering cost.

ZeroPoint claims DenseMem can create two-to-three times more effective CXL capacity through inline compression. No public customer case shows real compression ratios or savings.

Which AI memory startup can scale and has the strongest customer relationships?

d-Matrix and Axelera have the strongest evidence that they can build, deliver and support meaningful deployments. d-Matrix leads on deployment quality, Axelera on customer breadth.

d-Matrix has moved from chips to systems. Corsair combines 16 chiplets and more than 250 billion transistors, and the company now ships servers, racks and multi-rack clusters built with Supermicro, Arista and Broadcom. It runs its own Squadron cluster in an Arizona data center, and its NVIDIA roadmap plugs future Raptor chips into NVIDIA’s MGX supply chain. Parasail remains the clearest named production relationship in the field, with Gimlet Labs as a second.

Axelera has supported more than 600 customers and manufactures through TSMC and Samsung. Europa’s validation with Dell, Supermicro, HPE, Lenovo and others gives it the widest channel of any startup here. It has not said how many customers run large installations or place repeat orders.

MemryX has mass-production hardware and a growing relationship with Lenovo, while ZeroPoint has one semiconductor license. Panmnesia works with Meta on research, and EnCharge has not disclosed a production customer.

Which AI memory startup is gaining ground fastest right now?

d-Matrix has the strongest momentum, while XCENA is moving fastest among the companies that are still pre-volume.

d-Matrix has removed risk on several fronts in a few months: two acquisitions for systems and software, an NVIDIA partnership, and shipping clusters. The Information reported in July 2026 that d-Matrix was seeking new funding at a $5 billion valuation, up from $2 billion in its November 2025 Series C, and in October that NVIDIA plans to invest. Neither deal has been announced.

Axelera has followed a coherent path. It raised more than $250 million in February, launched Europa in September, signed its first AI factory supply contracts and added two non-executive directors in October. Its next test is turning breadth into large, repeat revenue. Axelera moving from the edge into the data center is why we point to the biggest opportunities in our AI chips market report.

XCENA took its $135 million Series B to work fast. It moved MX1 to a production lineup in August 2026, saw SK hynix show a CXL Hybrid Memory product built on MX1 in October, and hired a former Qualcomm and Intel executive as chief business officer to lead sales. Panmnesia co-authored a Nature Reviews Electrical Engineering paper with Meta in September 2026 on a CXL-based data-center design, a strong technical signal but not a purchase order.

RAAAM gained a heavyweight partner when GlobalFoundries announced joint GCRAM development and a test chip in September 2026. ZeroPoint’s momentum is quieter: it announced ZeroStream in July 2026 and keeps its Rebellions alliance and semiconductor license.

Is XCENA now ahead of Panmnesia in CXL memory?

In commercial terms, yes, narrowly. XCENA now has the stronger path to revenue, Panmnesia still has the more advanced CXL switching silicon, and ZeroPoint leads memory-compression IP. They sit in different layers, so one ranking only goes so far.

XCENA’s MX1 supports up to 2 TB of DDR5 memory per device and adds thousands of RISC-V cores that process data where it sits. In August 2026 XCENA turned its prototype into a production lineup, MX1 Compute and MX1 Expand. At OCP in October, SK hynix displayed a CXL Hybrid Memory product built on MX1 that adds SSD capacity behind DRAM, and the two companies say they are testing it in real AI inference settings. XCENA has still not published latency or throughput figures for its demos, nor disclosed a volume buyer.

Panmnesia has the strongest hardware evidence for advanced CXL fabrics. Its PCIe 6.4 and CXL 3.2 switch with port-based routing is sampling as pre-release silicon, and its controller IP now supports CXL 4.0 features. In July 2026 it presented silicon-proven switch and controller results at ISCA. Its September 2026 paper with Meta proposes running a whole data center like one chip over CXL, with up to 960 accelerators in one coherent domain. That shows Meta takes the idea seriously, but it is research, not a deployment.

ZeroPoint takes the least disruptive route. DenseMem is an IP block that sits inside a CXL controller or system-on-chip and compresses memory transparently. The company claims a two-to-three-times effective capacity gain and has one license with a major semiconductor maker.

Our judgment: XCENA leads on commercial traction because a memory giant is building a product on its chip. The gap is small, since that product is still in testing. Panmnesia leads on switching technology, and ZeroPoint on capital-light licensing.

Which AI memory startup has the strongest technical moat?

d-Matrix has the strongest broad technical moat because it now owns compute architecture, chiplets, memory hierarchy, networking, deployment software and production systems. ZeroPoint and RAAAM may hold stronger narrow IP moats, but their value depends on larger chipmakers adopting them.

Corsair is more than an accelerator with a different arithmetic unit. d-Matrix has built in-memory compute cores, custom numerical formats, chiplet links, its own JetStream network cards, compilers, runtime software and, through Wallaroo, orchestration tools. Raptor is designed to add 3D-stacked DRAM bonded directly to the compute die. Copying the product would require all of these layers.

Production experience deepens that moat. With Squadron running in its own Arizona data center and shipping to priority customers, d-Matrix sees model behavior, scheduling limits and failure modes that prototype-stage rivals cannot.

ZeroPoint’s compression IP and RAAAM’s patented GCRAM create narrower but licensable moats, and RAAAM’s GlobalFoundries program improves its odds of reaching real products. Panmnesia and XCENA remain technically differentiated, but neither has shown repeat deployments or switching costs yet.

Can NVIDIA or the memory giants erase these startup advantages?

They could, but for now they are more likely to absorb the leading startups than crush them. The winners so far are the startups that complement existing accelerators and memory rather than fight them.

d-Matrix is the clearest case. Its Raptor chips will plug into NVIDIA’s rack design through NVLink Fusion, with GPUs on prefill and d-Matrix on decode, and The Information reported that NVIDIA plans to invest. XCENA shows the same pattern with memory makers: SK hynix is building a product on MX1, and XCENA co-presented at OCP with Samsung. Every extra accelerator raises pressure on memory capacity, bandwidth and power, which helps all of these companies.

The threat is still real. NVIDIA can change its own memory hierarchy, add decode-specific hardware and fold specialized inference engines into future systems. Samsung, SK hynix and Micron can add compression, compute or new interfaces around memory. Hyperscalers can design custom silicon for their own workloads. Partnership can also become dependence: a startup inside NVIDIA’s rack sells on NVIDIA’s terms. Partnerships can turn into dependence, so we look at who the likely future winners are in our AI chips report.

Acquisition is a likely outcome too. Marvell paid an initial consideration of about $3.25 billion for Celestial AI and its photonic connectivity technology, and d-Matrix itself has turned buyer, absorbing GigaIO’s data-center business and Wallaroo.ai in 2026.

The most exposed startups are those selling an isolated performance gain without production know-how, software or embedded customer relationships.

How reliable is the evidence behind the AI memory startup race?

The public evidence is strong enough to name d-Matrix as the leader, but not strong enough to calculate market shares or crown a universal performance winner.

The best evidence is production availability, third-party testing, named deployments and real technology licenses. d-Matrix has all but a fully independent standardized benchmark; its newest OCP result is company-run. Axelera combines a large disclosed customer count with an independently reproduced edge benchmark and newly signed contracts. MemryX has one of the clearest independent product evaluations. ZeroPoint has a commercial license, though the counterparty is unnamed.

The weakest evidence is vendor-selected TOPS figures, memoranda of understanding, sales pipelines, reported but unconfirmed investments and projected savings without customer data. Axelera’s $1.5 billion pipeline and the reported NVIDIA stake in d-Matrix both fall in this group. Panmnesia’s Meta paper is credible research but not a paid rollout.

No company discloses enough on revenue, margins, backlog, retention, shipment volume or customer concentration. The ranking therefore measures competitive progress, not audited business scale.

Startup	Strongest available evidence	Main evidence gap	Overall confidence
d-Matrix	Full production, shipping clusters, named production deployment, partner workload testing and NVIDIA roadmap	Revenue, shipment quantities and neutral standardized benchmarks	Medium-high
Axelera AI	More than 600 customers, independently reproduced edge performance and signed AI factory contracts	Customer composition, deployment sizes and recurring revenue	Medium-high
MemryX	Mass-production hardware, independent hands-on evaluation and live Lenovo deployments	Customer count and commercial scale	Medium
ZeroPoint Technologies	License with a major semiconductor manufacturer	Customer identity, resulting product volumes and royalties	Medium
Panmnesia	Silicon-proven CXL products, pilot systems and joint research with Meta	Paid production deployment	Medium
XCENA	Production lineup, substantial funding and an SK hynix product built on MX1	Independent performance evidence and volume shipments	Medium
EnCharge AI	Commercial product launch and strong research foundations	Named customers, shipments and independent product testing	Medium-low
RAAAM Memory Technologies	Strategic semiconductor backing, licensing activity and a joint test-chip program with GlobalFoundries	High-volume customer silicon	Medium-low

Which AI chip memory startups are actually ahead?

d-Matrix is the AI chip memory startup ahead overall, and its lead has grown. It now combines shipping rack-scale systems, the most relevant data-center workload evidence, a multi-year NVIDIA roadmap and the strongest momentum. Axelera AI is the clear number two and leads at the edge, while MemryX holds third through mature hardware, live deployments and strong capital efficiency.

We give the most weight to production maturity and real deployments because chip startups often stall between a working prototype and a scalable product. Performance comes next when it reflects a useful workload. Customer quality, manufacturing readiness, defensibility and funding efficiency then decide whether a technical edge becomes a durable business.

d-Matrix does not win every metric. Axelera has far more disclosed customers. MemryX has stronger independent usability evidence. ZeroPoint may have the most capital-efficient commercial license. Panmnesia still has the most advanced CXL switch, while XCENA has moved narrowly ahead of it on commercial traction.

The ranking could change if Axelera’s AI factory contracts grow into large repeat orders, XCENA or Panmnesia lands a named hyperscaler rollout, or EnCharge secures a verified volume design win. d-Matrix could lose first place if Squadron shipments stay small or the reported funding and NVIDIA investment fail to close.

Rank	Startup	Why it holds this position
1	d-Matrix	The only core startup shipping multi-rack data-center inference systems, with a named commercial deployment, generative-AI workload evidence and an NVLink Fusion roadmap beside NVIDIA GPUs
2	Axelera AI	The clear edge-AI leader, with more than 600 customers, reproducible performance, a second chip shipping in Dell and Supermicro servers and funding comparable with d-Matrix
3	MemryX	Production hardware, excellent independent usability evidence, live Lenovo deployments and strong capital efficiency, but limited disclosure of customer and revenue scale
4	ZeroPoint Technologies	The strongest memory-compression specialist, supported by an actual semiconductor license and a low-disruption route into existing chips
5	XCENA	The best-funded computational-memory challenger after the top two, now with a production lineup and an SK hynix product built on its chip, but no disclosed volume buyer
6	Panmnesia	The leading independent CXL switch specialist, with advanced pre-release silicon and research backing from Meta, but no disclosed volume deployment
7	EnCharge AI	A technically credible analog in-memory contender with a launched 200-TOPS product, but no named volume customer or independent commercial benchmark
8	RAAAM Memory Technologies	Potentially valuable embedded-memory IP with GlobalFoundries now developing it, but customer products have not reached visible mass production