Oxmiq Raises $35 Million to Cut AI Chip Costs With Unified Design Push

San Francisco, July 1: Artificial intelligence startup Oxmiq has raised $35 million in fresh funding to develop a new chip architecture aimed at lowering the cost of building and operating AI systems, underscoring how the global race around artificial intelligence is increasingly shifting from software and models to the underlying hardware that powers them. The funding round, announced on July 1, puts renewed focus on one of the central questions confronting the AI industry today: how to make the next wave of computing infrastructure more affordable, more efficient and easier to scale.

Oxmiq’s pitch to investors is built around a simple but ambitious idea. Instead of relying on the traditional architecture in which graphics processing units, central processors and tensor engines are handled as separate components in an AI system, the company wants to collapse these into a single block of intellectual property that can be licensed to customers. In practical terms, the startup is trying to redesign the hardware foundation of AI so that model builders and data centre operators can reduce cost, complexity and power overhead at a time when demand for computing resources is exploding.

The company is led by Raja Koduri, one of the best-known names in the semiconductor world. Koduri previously served as chief architect at Intel and also held senior roles at AMD, where he was closely associated with graphics and accelerated computing. His return to the startup spotlight with a chip company focused on AI economics reflects a broader industry shift. The conversation around artificial intelligence is no longer only about which company has the most capable model or the smartest chatbot. It is increasingly about who can build the infrastructure that makes those systems economically viable at scale.

Oxmiq’s latest funding round takes its total capital raised to $60 million. The round included investors such as MediaTek and Pegatron Venture Capital, while Samsung Catalyst Fund and Fudomo led the financing. The money will be used to complete the company’s first batch of licensable intellectual property products and expand its engineering team as it moves toward commercialisation. That roadmap suggests Oxmiq does not want to compete solely as a chipmaker in the traditional sense. Instead, it appears to be positioning itself as a platform company whose designs and system architecture can be adopted by a broader set of players across the AI hardware ecosystem.

The timing of the fundraise is especially significant because AI computing costs have become one of the defining constraints of the current technology cycle. Training large models requires vast clusters of advanced chips, high-bandwidth memory, networking gear and specialised software stacks. Inference the process of actually running those models for users in production has become a major cost centre as well, especially as enterprises move beyond experimentation and start deploying AI tools across customer service, coding, design, search, analytics and automation. In this environment, every incremental gain in chip efficiency or system simplification matters.

That is where Oxmiq believes it has an opening. Today’s AI infrastructure is highly fragmented. Graphics processors handle parallel workloads, CPUs manage general-purpose tasks, and tensor engines or accelerators are often added to speed up matrix-heavy machine-learning operations. While this modular approach has enabled rapid innovation, it also creates complexity in system design, software optimisation and power management. Oxmiq wants to reduce those frictions by building a unified architecture that integrates multiple functions into a single licensable design block.

Koduri has framed the company’s ambition in sweeping terms, comparing Oxmiq’s aspiration to becoming “the Arm of this next era.” The reference is significant. Arm does not primarily make chips itself; it licenses chip designs and intellectual property that underpin a vast portion of the global smartphone market. If Oxmiq can create a similarly foundational role in AI infrastructure—by offering a standardised building block for next-generation AI systems it could potentially influence a much wider ecosystem than a single chip vendor typically would.

The company’s plans extend beyond a single chip block. Oxmiq also intends to develop a computing fabric that combines chiplets and memory into one package. Chiplets, which break a processor into smaller specialised components that can be assembled together, are increasingly seen as a key part of the future of semiconductor design. They allow companies to mix and match functions, improve manufacturing yields and accelerate customisation. In AI, where workloads vary widely between training, inference and edge deployments, chiplet-based architectures could offer a more flexible path to performance gains than monolithic chips alone.

By bringing chiplets and memory into a more integrated computing fabric, Oxmiq is effectively trying to address not just the chip itself but the broader system-level bottlenecks that make AI infrastructure expensive. Data movement between processors and memory remains one of the most significant constraints in AI performance and power efficiency. A system that reduces latency and simplifies communication between these components could lower both capital and operating costs for customers. In a market where electricity, cooling and server utilisation are becoming as important as raw compute performance, those savings could be commercially meaningful.

The startup’s strategy also reflects the growing fragmentation of the AI hardware market. Nvidia remains the dominant force in AI chips, particularly for training frontier models, but the scale of AI demand has opened room for a wide range of challengers and specialists. Established semiconductor companies, hyperscale cloud providers and venture-backed startups are all racing to design chips that are cheaper, more efficient or better tailored to specific workloads. Some are targeting inference, others edge AI, others custom accelerators for data centres. Oxmiq appears to be pursuing a slightly different path: not simply selling another chip, but rethinking how the architecture of AI systems is packaged and licensed.

That distinction matters because the economics of AI hardware are brutal. Developing a cutting-edge chip can cost hundreds of millions of dollars and take years of design work, software support and manufacturing coordination before it ever reaches customers. Startups face a difficult choice: either raise enormous sums to build a full chip business, or find a narrower point of leverage in the stack. Oxmiq’s licensing approach may offer a way to do the latter. By focusing on intellectual property blocks and system architecture, the company could potentially influence multiple products and customers without shouldering the full burden of manufacturing and distribution at every stage.

Still, the challenges are formidable. The semiconductor industry is crowded with incumbents, and AI hardware buyers are notoriously demanding. Customers need proof not only that a new architecture is technically elegant, but that it delivers measurable performance gains, integrates with existing software frameworks and can be supported over long product cycles. Oxmiq will have to convince potential partners that its unified design can fit into real-world AI deployments where reliability, compatibility and developer tooling are as important as raw efficiency.

There is also the question of timing. The AI market is moving fast, and hardware cycles do not. Startups in this space often face a race against both technological obsolescence and changing customer expectations. If model architectures evolve in ways that reduce dependence on certain hardware assumptions, or if hyperscalers build more of their own in-house solutions, the window for a third-party architecture provider can narrow quickly. At the same time, however, the scale of global demand for AI infrastructure is so large that even a modest foothold in the market could prove valuable.

The Oxmiq funding round lands at a moment when the AI industry is wrestling with an uncomfortable reality: the current pace of spending may not be sustainable unless the cost of compute comes down. Data-centre operators, cloud companies and model developers are pouring billions into infrastructure, but investors are increasingly asking whether the economics of generative AI will support those investments over the long term. Startups that can reduce the cost per token, per inference request or per training run therefore occupy a strategically important position. They are not merely component suppliers; they are potential enablers of the entire business model.

In that sense, Oxmiq’s announcement fits into a much larger story about the industrialisation of AI. The first public phase of the generative AI boom was dominated by dazzling demos, conversational agents and viral image generators. The second phase is increasingly defined by less glamorous but more consequential questions: how to power these systems, how to finance the infrastructure behind them, how to secure enough chips, and how to keep the economics from spiralling out of control. Hardware startups like Oxmiq sit at the heart of that transition.

The involvement of investors such as MediaTek and Samsung Catalyst Fund also hints at the strategic interest surrounding Oxmiq’s technology. Large semiconductor and electronics companies are watching closely for architectures that could reshape the economics of AI systems, particularly if those designs can be licensed or adapted across multiple markets. For investors, backing a company at the architecture layer offers exposure to AI’s long-term infrastructure buildout without necessarily betting on a single application or model provider.

Koduri has indicated that Oxmiq also plans to enter the custom chip market, where companies such as Broadcom, Marvell and MediaTek already compete to design silicon tailored to the needs of specific customers. That ambition adds another layer to the startup’s strategy. Custom chips are becoming increasingly attractive for hyperscalers and large enterprises that want to optimise performance and cost for their own AI workloads. If Oxmiq can pair a licensable architecture with custom silicon capabilities, it could position itself as both a design enabler and a product partner in the broader AI hardware supply chain.

The broader market context makes such ambitions plausible, even if difficult. Cloud providers and large technology firms are no longer content to buy only off-the-shelf hardware. They want more control over cost, power efficiency and workload optimisation. That has created new openings for chip startups, especially those that can offer differentiated intellectual property rather than compete head-on with Nvidia in the most crowded parts of the market. Oxmiq’s focus on unified design, chiplets and system level integration appears tailored to this moment.

Whether the company can deliver on its promise remains uncertain, but its fundraise is an important marker of where capital is flowing in the AI era. Investors are not only backing model developers and software startups; they are increasingly placing bets on the plumbing of artificial intelligence itself. Chips, memory systems, packaging technologies and compute fabrics have become some of the most strategic layers of the tech stack. Oxmiq’s $35 million round is modest by the standards of hyperscale AI spending, but it reflects confidence that there is still room for new ideas in a market dominated by giants.

For the technology industry, the significance of the announcement lies less in the size of the cheque and more in what it reveals about the next frontier of competition. AI’s future will not be determined solely by who writes the best model weights or launches the flashiest chatbot. It will also be shaped by who can redesign the economics of compute, lower the cost of infrastructure and make advanced AI more practical to deploy at scale. Oxmiq is betting that the answer lies in unifying the hardware building blocks beneath the AI boom.

If that bet pays off, the startup could carve out an influential role in the evolving architecture of artificial intelligence. If it fails, it will still have illustrated one of the defining truths of the current moment: that in the AI economy, the battle for advantage is increasingly being fought not only in code and data, but in silicon, system design and the cost of every watt and every calculation. For now, Oxmiq’s funding round is another sign that the race to shape AI’s infrastructure is accelerating and that some of the most consequential innovations may come from the companies trying to make the technology cheaper to run rather than simply more powerful to use.