Everspin, MaxLinear Sign an MOU. That Is Not the Same as a Product Deal.

Everspin Technologies and MaxLinear signed a memorandum of understanding today to evaluate combining MRAM with storage compression for AI servers. Everspin's stock jumped 14%. MaxLinear added 9%. The market heard two companies entering the AI infrastructure war together.

What they actually announced is a framework for testing. An MOU - a non-binding letter of intent - is a marketing asset, not a commercial contract. There is no product shipped, no revenue timeline, no customer named, and no price attached.

An MOU is what companies sign when they want the press release without the engineering commitment.

Let's look at what's underneath the headline.

The partnership targets "metadata management, log data, write buffering and cache functions" in AI servers. That is a narrow set of support-layer tasks, not the memory hierarchy where AI workloads actually spend their budget. The bottleneck that defines AI server economics is the HBM (high-bandwidth memory) attached directly to the GPU - the DRAM that feeds tokens through the key-value cache during inference. MRAM, the non-volatile magnetic memory Everspin makes, operates at speeds and densities far below HBM. It cannot replace the memory tier that matters most to AI inference performance.

The roles these two companies are positioning themselves for - metadata, logs, write buffers - are the leftover workloads after HBM, DDR, and even SCM (solid-state cache memory) have done their job. That is not nothing. But it is not what drives the AI server build decision. It's the equivalent of designing a better glovebox for a car and calling it a breakthrough in transportation.

Then there's the scale problem. Everspin reported $14.9 million in Q1 2026 revenue, which annualizes to roughly $59 million. Its total equity is $70 million. The stock has been up 66% year-to-date, running on the assumption that MRAM is the "next big thing in AI," as one Yahoo Finance headline put it last month.

Everspin is a $70-million-equity company selling a niche memory technology to casinos, automotive, and industrial customers. The fact that Kerrisdale Capital - an actual semiconductor research shop - published a report three months ago calling Everspin "not an AI infrastructure winner" should have been the story. Instead the market is trading on an MOU.

Kerrisdale's analysis, published May 11, was blunt: Everspin is a niche industrial memory supplier whose largest end markets include casino gaming systems, not hyperscale data centers. The company has been growing revenue in the mid-teens percentage range - from $10.6 million in Q2 2024 to $14.9 million in Q1 2026 - which is respectable for a specialty vendor but does not suggest it is on the cusp of displacing anyone in the AI memory stack.

The financials don't suggest a company building infrastructure-scale products. Everspin's operating margin is -12.8%, its ROIC is -11.1%, and free cash flow has deteriorated to a negative $3 million trailing twelve months. Capital expenditure runs $12 million annually. The company has $40 million in cash and essentially no debt, which means it's not leveraged into trouble - but it also means it lacks the balance sheet heft to fund the multi-hundred-million-dollar fab expansions that memory scale requires.

MaxLinear is the more interesting half of this pairing. Revenue grew 50% year-over-year, and the company has been actively pitching its Panther storage accelerator platform for AI data movement, compression, and encryption. MaxLinear showcased Panther at the Future of Memory and Storage conference this week, demonstrating hardware-accelerated compression claiming up to 3× compression ratios on AI data. The company's stock is up 285% year-to-date - a violent move that has clearly priced in far more AI optimism than current earnings can support. MaxLinear's operating margin sits at -19.3%, ROIC at -15.6%. It is spending $15.8 million on capex and carries $338 million in debt against $65 million in cash. The stock moved from a 52-week low of $12.77 to a high of $128.30. That kind of move precedes disappointment when the revenue behind it is not yet proven at hyperscale.

The global MRAM market itself is under $1 billion today. Even the most bullish forecasts put it at roughly $3.9 billion by 2033, growing at an 18.7% CAGR - which is solid for a specialty memory but nowhere near the addressable universe of the HBM market, which is already in the tens of billions and growing faster. Everspin's share of an under-$1-billion market, against competitors including Toshiba's embedded MRAM products and potential future entries from the major foundries, is a narrow moat at best.

What the MOU actually tells us is that both companies need the AI narrative. Everspin, whose core customers are casino and industrial OEMs, needs to convince the market that MRAM belongs in the AI stack. MaxLinear, whose stock has already run up nearly 300% on the AI storage story, needs partnerships that look like validation to justify its valuation. Together, they can produce a press release that reads bigger than either company's current revenue.

The cross-currents here are clear. MRAM has legitimate technical properties - non-volatility, endurance, radiation tolerance - that make it useful in specific embedded and industrial applications. MaxLinear's Panther compression platform may address a real pain point as AI agent workloads expand metadata and cache demands. Both companies are growing revenue.

Directionally, however, the gap between today's announcement and a commercial AI server product using Everspin MRAM with MaxLinear compression is large enough to cross an ocean, and neither company has shown the engineering scale to swim it.

The engineering questions remain unanswered. Can Everspin's MRAM meet the latency and density requirements of even the secondary memory tiers in a next-generation AI server? Has it been validated against any actual server platform, or does "evaluation" mean they haven't tried yet? Does MaxLinear's compression actually deliver the claimed 3× ratio on the specific MRAM-backed workloads described in the MOU, or were those demos run on conventional NVMe?

Any astute semiconductor investor would have looked at the Kerrisdale report, the $60 million revenue run rate, the -13% operating margin, and the non-binding nature of an MOU before buying into the AI server headline. The stock move today - 14% for Everspin, 9% for MaxLinear - is pricing a commercial outcome that has not been engineered, validated, or contracted.

The thesis for investors is simpler than the press release suggests. Until Everspin demonstrates that MRAM is deployed in production AI server architectures at meaningful volume, and until MaxLinear shows that Panther compression is purchasing through hyperscale procurement channels, these two stocks are trading on narrative, not infrastructure. The MOU is a step, but it is not the arrival.