AMD faces a formidable challenge in its rivalry with NVIDIA, but recent developments signal the chipmaker’s determination to compete not just on hardware, but across the entire AI ecosystem.

Two separate announcements this week—AMD’s acquisition of inferencing software startup MK1 and a $6 million funding round for GPU portability framework developer Spectral—highlight the critical importance of software infrastructure in the GPU wars.

“AMD is entering a new era of growth fueled by our leadership technology roadmaps and accelerating AI momentum,” said Dr. Lisa Su, AMD chair and CEO.

“With the broadest portfolio of products and our deepening strategic partnerships, AMD is uniquely positioned to lead the next generation of high-performance and AI computing,”

“We see a tremendous opportunity ahead to deliver sustainable, industry-leading growth. We have never been better positioned.” she said.

AMD Unveils Strategy to Lead the $1 Trillion Compute Market 

At its Financial Analyst Day, AMD (NASDAQ: AMD) showcased its long-term strategy, leadership products and technology IP, underscoring the company’s momentum in driving accelerated growth and delivering long-term shareholder value.

The NVIDIA Moat: More Than Silicon

While AMD has made significant strides in datacenter revenue—posting $12.6 billion in datacenter sales for 2024, a 94% year-over-year increase according to company filings—the competition extends far beyond chip specifications.

NVIDIA maintains an overwhelming market position, holding approximately 90% of the discrete GPU market in Q3 2024, according to Jon Peddie Research data. By comparison, AMD’s market share in gaming GPUs fell to just 10% in Q3 2024 before recovering slightly to 17% in Q4.

The disparity is even more stark in AI accelerators. Despite AMD generating $5 billion in Instinct GPU revenue in 2024, analysts estimate this represents only 4-5% of the AI accelerator market—a figure that has remained stagnant despite year-over-year growth

NVIDIA’s advantage lies not just in its hardware, but in CUDA—its proprietary software framework that has become the de facto standard for GPU computing.

Thousands of applications and an entire generation of developers have been trained on CUDA, creating a powerful lock-in effect that AMD must overcome.

Breaking the CUDA Lock

Enter Spectral, a San Francisco-based company that on November 10 announced a $6 million seed round led by Costanoa Ventures.

Founded in 2018, Spectral has developed SCALE, a toolchain that compiles CUDA code to run natively on AMD chips without modification or performance penalties.

Unlike translation layers or emulators, SCALE functions as a superset of both CUDA and AMD’s ROCm framework, producing independent runtimes optimised for AMD hardware.

The timing is strategic. Multiple cloud providers now offer AMD chips as alternatives to NVIDIA, and GPU cloud startup TensorWave has built its entire business model around AMD architecture.

Spectral plans to expand SCALE support to other hardware platforms, including Intel GPUs and hyperscaler-designed chips such as AWS’s Trainium and Google’s TPUs.

Optimising for AMD’s Architecture

AMD’s acquisition of MK1, announced this week for an undisclosed sum, addresses a different piece of the puzzle: inference optimisation.

Founded in 2023 by Paul Merolla, a former Neuralink co-founder, MK1 focuses on making AI inference more efficient on AMD hardware. The acquisition appears tied to AMD’s technical advantages in certain areas, particularly memory capacity.

The MI355X offers 288 GB of on-chip memory compared to 180 GB for NVIDIA’s GB200, allowing models to run on fewer GPUs—a significant cost consideration for large deployments.

MK1 had previously partnered with TensorWave and posted benchmark results touting “DeepSeek-style optimisations” on AMD chips, suggesting the startup had identified architectural opportunities specific to AMD’s design choices.

A Pattern of Strategic Acquisitions

The MK1 acquisition fits within AMD’s broader acquisition strategy aimed at building competitive advantages against NVIDIA. Recent deals include:

• Silo AI ($665 million, 2024): European AI lab acquisition
• ZT Systems ($4.9 billion, 2024): Rack-scale systems expertise
• Enosemi (undisclosed): Silicon photonics technology
• Brium (undisclosed): Compiler technology

According to SEC filings, AMD made small acquisitions totaling $36 million over nine months, likely accounting for MK1, Enosemi, and Brium combined.

The ZT Systems acquisition proves particularly significant. Modern GPU clouds and “neoclouds” require rack-scale, row-scale, or even datacenter-scale deployment capabilities—not individual server sales.

AMD’s forthcoming MI400 series, scheduled for release next year, will incorporate this rack-scale expertise, bringing the company closer to offering complete systems comparable to NVIDIA’s integrated solutions.

Ecosystem Momentum Builds

AMD can point to notable deployments: Meta exclusively used MI300X to serve its Llama 405B frontier model, while Microsoft employs MI300X to power multiple GPT-4-based Copilot services.

In high-performance computing, AMD now powers five of the world’s 10 fastest supercomputers, including El Capitan, currently ranked number one.

Moreover, AMD achieved a historic milestone in Q4 2024, outselling Intel in datacenter revenue for the first time, with AMD’s datacenter segment generating $3.86 billion compared to Intel’s $3.4 billion.

Yet the challenge remains substantial. Even as AMD’s datacenter GPU business grows, analysts project relatively flat first-half 2025 revenue compared to second-half 2024, with acceleration expected only after the MI350 launch later in the year.

The Long Game

The investments in MK1 and Spectral’s funding round represent essential but incremental progress. AMD must not only produce competitive hardware but also cultivate an entire ecosystem—including developer tools, optimised software frameworks, trained developers, and customer confidence.

With the AI accelerator market projected to reach $1 trillion by 2030, according to AMD’s own estimates, the stakes justify the investment.

Whether software acquisitions and framework development can meaningfully erode NVIDIA’s overwhelming market position remains the defining question for AMD’s AI ambitions.

Industry observers note that while competition benefits customers through innovation and potentially lower prices, AMD faces the dual challenge of keeping pace with NVIDIA’s rapid innovation cycle while simultaneously building the supporting infrastructure that makes its chips truly competitive

The race is far from over, but AMD’s recent moves suggest a company that understands hardware alone won’t win this battle.