In November 2025, Majestic Labs closed a $90M Series A led by Bow Wave Capital, which it announced alongside its emergence from stealth.

Prior to the Series A, Lux Capital led a $10M seed round. Additional participants across the two rounds include SBI, Upfront Ventures, Grove Ventures, Hetz Ventures, QP Ventures, Aidenlair Global, and TAL Ventures.

Total disclosed funding stands at over $100M.

Majestic Labs builds Prometheus, a rack-scale AI server organized around a single large shared memory pool instead of a cluster of GPUs with separate local memory.

Prometheus is built to address what the company calls the memory wall: modern AI processors are fast, but they spend a disproportionate amount of time waiting for data from memory. In a conventional GPU cluster, memory is fragmented across tiers: each GPU has its own high-bandwidth memory, other GPUs' memory is accessible but slower, and host memory is slower still. That forces developers and infrastructure teams to manage where model weights, KV cache, and context data physically live. Prometheus presents the memory space as one flat, uniform pool.

A single Prometheus system can be configured with between one and twelve compute chips and between 8 TB and 128 TB of memory. The system uses standard LPDDR rather than the scarce and expensive HBM used in high-end GPU systems. Majestic Labs' custom memory interface chiplets handle the bandwidth and flow-control work needed to make that pool fast enough for AI inference.

On the compute side, the system uses Ignite, Majestic Labs' processor architecture, which combines ARM cores with RISC-V vector and tensor cores. Majestic Labs has licensed accelerator IP from a third party rather than building every compute primitive in-house, concentrating its own engineering on the memory interface and system architecture, which it sees as the primary bottleneck.

For users, the workflow is designed to stay familiar. Prometheus supports PyTorch, vLLM, and Triton without code rewrites, and the software stack can already take Hugging Face models and lower them to executable code for a simulation of the server. Majestic Labs also envisions hybrid deployments where customers use Nvidia hardware for prefill and Prometheus for decode, letting the product enter accounts as a specialized memory-dense tier instead of requiring a full infrastructure replacement on day one.

Majestic Labs sells B2B as a vertically integrated AI infrastructure vendor. Its go-to-market is direct enterprise sales to hyperscalers, neoclouds, high-frequency trading firms, and large enterprises with memory-heavy AI workloads, with long evaluation cycles, proof-of-concept benchmarking, and pilot deployments preceding any volume purchase.

Its monetization centers on system-level value capture rather than component pricing. Majestic Labs argues that one Prometheus server can replace multiple racks of conventional GPU infrastructure, reducing power draw, cooling burden, physical footprint, and the cost of overbuying compute to obtain enough attached memory. That framing supports pricing at a premium to ordinary server hardware while preserving a total cost of ownership argument.

Using standard LPDDR instead of HBM is a deliberate cost-structure decision. It reduces bill-of-materials pressure and supply chain exposure at a moment when HBM allocation is a strategic constraint across the industry, and gives Majestic Labs more room to price on value rather than component cost.

The company is also building a full software stack, compilers, runtime, and model-serving orchestration, creating a path toward higher-margin software and support revenue attached to hardware deployments. The longer-term question is whether Majestic Labs remains a hardware appliance vendor or uses its memory-first differentiation as a wedge into a managed inference or cloud service model, the path Cerebras has increasingly pursued.

Majestic Labs enters a market where the basis of competition has shifted from chip versus chip to system versus system. Nvidia sets the reference architecture, while a growing set of inference specialists targets specific workload niches with different designs.

Nvidia is the default benchmark for any AI infrastructure vendor, and Majestic Labs frames its product positioning against Nvidia rack economics. Nvidia's Vera Rubin NVL72 is a rack-scale platform combining 72 Rubin GPUs, over 20 TB of HBM4, tens of terabytes of LPDDR5X on paired CPUs, sixth-generation NVLink, and an established ecosystem of integrators and software partners.

For Majestic Labs, the challenge is that Nvidia sells more than silicon: procurement trust, CUDA ecosystem inertia, and a roadmap that keeps extending memory, interconnect, and inference performance within the incumbent stack. A buyer comparing Prometheus to a Vera Rubin NVL72 deployment is not comparing specs in isolation; they are comparing a novel architecture with a vendor that already has budget line items and integrator relationships. AMD's Instinct roadmap adds a second mainstream option for buyers that want non-Nvidia GPU infrastructure without adopting a startup architecture, creating an alternative path for fence-sitters.

Groq and Cerebras are the most direct inference-specialist competitors. Groq has commercialized its purpose-built LPU primarily through GroqCloud, which serves millions of developers across thirteen data centers and processes trillions of tokens weekly, with a June 2026 growth round funding expansion toward 200 MW of capacity by 2027. Relative to Majestic Labs, Groq's advantage is a cloud service that is available today, with a low-friction developer entry point that does not require customers to wait for 2027 hardware.

Cerebras is the closest conceptual rival on the thesis that decode and long-context inference are memory-bandwidth constrained. Cerebras has positioned its CS-3 systems as the decode layer of disaggregated inference stacks, with AWS-linked commercialization paths and marketplace access, and Sacra estimates Cerebras reached $510M in revenue in 2025. Both companies are selling the view that generic GPUs are poorly matched to frontier inference workloads, but Cerebras already has a visible cloud and API distribution layer, while Majestic Labs is earlier and more hardware-system centric.

d-Matrix's Corsair platform entered full production in June 2026, targeting hyperscalers, neoclouds, and frontier labs with heterogeneous disaggregated compute that pairs its accelerators with existing GPU fleets. That complement-rather-than-replace positioning contrasts with Majestic Labs' single-system pitch, and buyers often adopt complements before substitutes.

Etched is vertically co-designing chips, racks, software, and manufacturing for frontier inference clusters, with first racks shipping in summer 2026 and over $1B in reported demand. SambaNova competes less as a chip rival and more as a procurement-model rival, offering turnkey managed inference deployed inside the customer's own data center, a lower operational burden for enterprise buyers that want AI infrastructure without becoming hardware operators. Tenstorrent adds pricing and ecosystem pressure from the open-hardware direction, with Galaxy servers and vLLM-compatible serving for buyers that want non-Nvidia optionality without betting on a radically new system architecture.

Majestic Labs' expansion logic starts from a narrow but real wedge, memory-bound inference for very large models, and extends into adjacent products, software, and customer segments as the product matures and demand for private AI infrastructure grows.

The most immediate expansion vector is broadening Prometheus from a single system into a family of SKUs tuned for different workload profiles, including high-end frontier-model servers, lower-cost enterprise appliances, and configurations optimized for specific use cases like long-context inference, graph neural networks, video generation, or tabular models.

A second expansion surface is software. As compilers, runtime, and model-serving orchestration layers mature, Majestic Labs could monetize them through licensing or support contracts, shifting from a pure hardware revenue model toward a bundled hardware-plus-software platform similar to Rebellions and Cerebras, which pair hardware with SDK and inference API layers.

Majestic Labs' near-term beachhead is frontier AI labs, hyperscalers, and neoclouds, but the larger addressable market lies in regulated and data-rich enterprises such as financial services, healthcare, government, and industrials, where data locality, power efficiency, and TCO matter more than the fastest GPU benchmark.

The mid-tier enterprise and regional cloud segment is another plausible expansion target because Majestic Labs' pitch that one server can replace multiple racks of conventional infrastructure is more relevant for organizations that cannot support hyperscaler-style multi-rack buildouts on power, space, or operational complexity grounds. Sovereign AI programs are a third customer category: governments and national champions investing in domestic AI supercomputing clusters often want advanced capabilities with a smaller footprint and more local control than hyperscaler-dependent deployments, and Majestic Labs' compact, power-efficient architecture maps to that requirement.

Majestic Labs' leadership has ties across Google, Meta, and hyperscaler ecosystems in APAC, making Japan and broader Asia-Pacific a logical early international expansion region, particularly where power-efficient private AI infrastructure is attractive to large industrial groups.

Europe is a second geography given rising sovereign AI infrastructure commitments and the practical appeal of rack-scale-in-a-box deployments in regions where power availability and real-estate constraints favor compact systems. In the United States, TSMC's expanding Arizona capacity and domestic AI infrastructure policy initiatives create another angle for Majestic Labs to position Prometheus as a power-efficient alternative to large GPU estates for defense-adjacent and enterprise programs.

Hardware timing risk: Majestic Labs' compute and memory-interface chips are targeting tape-out in 2026 with lead-customer shipments in 2027, which compresses the first-silicon-to-production transition in a foundry and advanced-packaging environment under AI-driven demand pressure, where delays between architectural targets and reliable volume delivery have historically been common across AI chip startups.

Incumbent convergence: Nvidia's Vera Rubin NVL72 already integrates tens of terabytes of LPDDR5X alongside HBM4 in a rack-scale system, Cerebras is productizing disaggregated decode inference with AWS distribution, and d-Matrix is shipping heterogeneous accelerators that complement rather than replace GPU fleets, so the memory-bound inference bottleneck Majestic Labs is targeting is already being addressed from multiple directions, which narrows the window for its architecture to emerge as the preferred commercial solution.

Workload scope narrowing: If customers solve memory-bound inference economics through quantization, improved KV-cache techniques, model efficiency advances, or software-defined memory expansion tools, Majestic Labs' addressable market could contract to the most memory-hungry workloads rather than broadening into a wider replacement for conventional GPU infrastructure.