RadixArk Lands Nvidia, AMD Backing For AI Software Push That Challenges Enterprise Tools

The round, led by Accel and Spark Capital, drew participation from Nvidia's NVentures corporate venture fund, AMD directly, Databricks Inc., and Broadcom Inc. Chief Executive Hock Tan, among others. The Wall Street Journal reported the $400 million valuation on May 5, 2026. The breadth of the syndicate, spanning chip design, cloud data infrastructure, and enterprise software, reflects a consensus among hardware and platform vendors that the bottleneck in AI deployment has shifted from model capability to inference efficiency and developer tooling .

RadixArk builds commercial products on two open-source foundations: SGLang, an inference framework its team helped create before the company's founding, and Miles, a training optimization library the company open-sourced in November 2025. The company plans to sell managed infrastructure and cloud hosting services built on these tools, according to the Wall Street Journal .

The implication for institutional capital is direct. At a $400 million post-money valuation on $100 million raised, investors are pricing RadixArk at a 4x price-to-capital ratio at seed, a premium that reflects the scarcity value of teams with deep, production-proven open-source credibility. The question is whether the commercial monetization layer can scale fast enough to justify what will likely be a much higher valuation at Series A.

SGLang's 400,000-GPU Footprint Is the Moat Investors Are Buying

The number that anchors the RadixArk bull case is not its valuation. It is the scale of SGLang's existing deployment. According to RadixArk, SGLang currently powers AI clusters containing more than 400,000 graphics cards in aggregate . That is not a pilot program. That is production infrastructure at hyperscale.

SGLang's core technical contribution is KV cache reuse. Large language models generate substantial temporary data when processing prompts, data stored in what is called a KV cache. Standard implementations clear that cache after every prompt. SGLang enables models to reuse portions of the cache across prompts, reducing the computational overhead of regenerating data from scratch on each request and accelerating response times in the process .

The framework compounds this efficiency through two additional mechanisms. Speculative decoding offloads portions of a prompt's processing to a lighter, less hardware-intensive model running in parallel with the primary LLM. Separately, SGLang can distribute the calculations involved in processing a single prompt across chips with different architectures, a capability that carries particular relevance given AMD's participation in the round. AMD has been pushing its ROCm software stack as a viable alternative to Nvidia's CUDA ecosystem. A commercial SGLang offering that is architecture-agnostic would give AMD a toolchain argument to deploy at enterprise accounts currently locked into Nvidia infrastructure.

Our view: Nvidia's investment in RadixArk is not altruistic. SGLang's ability to run across heterogeneous chip architectures is a double-edged instrument. Nvidia gains distribution and ecosystem lock-in if SGLang defaults to optimizing for its hardware. But if SGLang becomes a true abstraction layer, it could erode the switching costs that give CUDA its durability. Nvidia is buying optionality and, more importantly, buying influence over a project that already runs on 400,000 of its own GPUs.

Miles Targets the Trillion-Parameter Compression Problem

The second pillar of RadixArk's technology stack addresses a different constraint: training cost and memory footprint. Miles, open-sourced by the company in November 2025, can compress large language models with one trillion parameters into a format that fits within the memory of a single high-end graphics card . That capability, if it holds at commercial scale, directly reduces the hardware expenditure required to train and iterate on frontier-class models.

Miles also includes a component called MrlX, described as an asynchronous co-evolutionary framework. MrlX places multiple AI agents into a shared simulated environment and trains them concurrently, allowing the agents to improve their reasoning by observing and learning from one another . The technique is a variant of multi-agent reinforcement learning and is particularly relevant for teams building reasoning-heavy LLMs, the same category of models that have driven recent benchmark gains across the industry.

The competitive context matters here. Subquadratic, another AI infrastructure startup that launched the same day, May 5, 2026, with $29 million in seed funding, is attacking a related problem from a different angle. Subquadratic's SubQ model uses a sparse attention architecture to deliver a 12 million-token context window while claiming to be more than 50 times faster and 50 times less expensive than leading frontier models at 1 million tokens. At its full context window, the company claims compute requirements fall by nearly 1,000 times compared to other frontier models, with a score of 95% accuracy on the RULER 128K benchmark at a cost of $8, versus approximately $2,600 for Claude Opus at comparable accuracy .

These are not competing companies in the direct sense. But they are competing for the same institutional insight: that the frontier model arms race is becoming secondary to the infrastructure efficiency race. Capital is now pricing the picks-and-shovels layer of AI, not just the models themselves.

The Hock Tan Participation Signals a Networking and Inference Convergence Play

Broadcom CEO Hock Tan's personal participation in the RadixArk round is the detail most institutional investors should examine first. Tan's Broadcom has been aggressively positioning its custom AI accelerator (XPU) business as an alternative to both Nvidia GPUs and AMD Instinct chips for hyperscaler inference workloads. Broadcom does not comment on private investment terms, and the financial size of Tan's personal commitment was not disclosed .

The strategic logic, though, is readable. If SGLang can distribute inference calculations across heterogeneous chip architectures, and if Miles can compress model weights to fit on a single card, then a customer running Broadcom XPUs alongside Nvidia or AMD GPUs gains a software layer that maximizes utilization across the entire fleet. For Broadcom, that is a commercial argument for XPU adoption that is currently difficult to make without a framework like SGLang to tie the stack together.

The participation of Databricks is also notable. Databricks has been building toward a full-stack AI development platform, and its investment in RadixArk suggests Miles and SGLang are candidates for integration or distribution through the Databricks ecosystem. Databricks has not announced any formal product partnership with RadixArk. Terms were not public.

AI Tooling Is Becoming the New Infrastructure Battleground

The RadixArk round does not exist in isolation. It sits inside a broader reallocation of AI investment from foundation model training to the application and tooling layers that convert compute into deployable, cost-efficient products. Harvey, the legal AI startup, reported in May 2026 that its users now run more than 700,000 agent-powered tasks per day, and that hours spent in its software per user per month have risen 75% over the prior four months, growth the company attributes largely to agent adoption . Harvey's CEO Winston Weinberg told Business Insider that agent usage is "exploding" among customers even as the company is in early days of teaching firms how to deploy them .

The pattern is consistent: the monetization of AI is increasingly happening not at the model layer but at the workflow and tooling layer. RadixArk's plan to build managed infrastructure and cloud hosting products on top of SGLang and Miles is precisely this thesis. Developers who already use the open-source versions become the natural commercial conversion funnel. The open-source installed base, anchored by SGLang's 400,000-GPU deployment footprint, provides distribution that most enterprise software startups spend years and hundreds of millions of dollars trying to build.

The outstanding question is pricing power. Open-source adoption creates scale but not necessarily margin. RadixArk's commercial model, described only as managed infrastructure and tooling with cloud hosting capability, was not detailed in terms of pricing structure or contract size . The company will need to demonstrate that enterprise customers will pay a material premium over self-managed open-source deployment.

The Plocamium View

The RadixArk round is best understood as the market paying a premium for open-source distribution that has already escaped the lab. The $400 million valuation on a company with no disclosed revenue is not irrational if you accept one premise: that SGLang's 400,000-GPU installed base is a demand signal, not a vanity metric. That footprint means RadixArk is already embedded in the production workflows of organizations making real infrastructure decisions. Converting a fraction of that installed base into paying commercial customers is a more executable path than greenfield enterprise sales.

The more interesting second-order question is what this round means for the competitive dynamics between Nvidia and AMD. Both companies are now co-investors in a framework that abstracts away hardware-specific optimization. Historically, Nvidia's CUDA ecosystem has generated durable competitive advantage precisely because it is hardware-specific. SGLang's heterogeneous dispatch capability introduces a wedge. If it matures, it becomes a negotiating tool for hyperscalers and enterprises to extract better pricing from both chipmakers.

Plocamium's thesis: the AI tooling layer is undergoing a valuation re-rating in 2026 that mirrors what the cloud middleware layer experienced between 2012 and 2016, when companies like HashiCorp, Confluent, and Elastic built durable businesses on open-source foundations with commercial distribution layers on top. RadixArk's comparable set is not today's AI model companies. It is that generation of infrastructure software businesses. The multiples those companies achieved at exit, typically between 10x and 20x revenue at IPO for best-in-class performers, represent the ceiling RadixArk investors are underwriting. The floor is whether the commercial monetization closes before a hyperscaler decides to commoditize SGLang's core KV cache and speculative decoding features natively. That risk is real, and the 18-to-24-month window before a Series A will determine which scenario dominates.

Investors with exposure to Nvidia and AMD should watch RadixArk's commercial traction not for the startup's own valuation trajectory, but as a leading indicator of whether the inference efficiency layer is consolidating around independent tooling vendors or folding back into the chip vendors' own software stacks.

The Bottom Line

RadixArk's $100 million seed round, backed by the two largest GPU vendors plus a Broadcom CEO writing personal checks, is the clearest signal yet that the infrastructure layer of AI is being contested at the software level, not just the silicon level. The company enters the commercial market with a production-scale open-source installed base, a technically credible compression and inference stack, and a syndicate of investors with direct commercial incentive to see it succeed. The next 12 months will test whether that open-source gravity translates into enterprise revenue. If it does, this round will look like the cheapest entry point in the company's history. If the hyperscalers internalize the same capabilities, the $400 million valuation will look like a product of investor exuberance at the peak of an infrastructure cycle. The answer will arrive before the next round closes.