InvestmentGuru flags $PENG $MRAM $MXL

AI infrastructure is getting wider. That is the real story here. Nvidia still gets the attention, but money is now spilling into the less obvious parts of the stack — memory expansion, optical interconnects, and specialized embedded memory. That is why PENG, MRAM, and MXL keep showing up together in trader chatter. But turns out they are not one theme. They are three separate bets wearing the same AI label. ### What does PENG actually sell? Penguin Solutions is basically a systems company. It builds and integrates AI and high-performance computing infrastructure, with a growing pitch around memory-heavy inference systems. In its April 1 fiscal Q2 2026 report, the company raised full-year net sales and EPS outlook, even though quarterly sales were down year over year, because management said demand was improving in AI/HPC and memory. The interesting detail is not “AI exposure.” It is that Penguin is trying to win where memory becomes the bottleneck, including CXL-based server designs aimed at inference workloads. (ir.penguinsolutions.com) ### Why does memory matter so much? Training gets the headlines, but inference is where memory pressure gets ugly. Large models need fast access to huge working sets, and that makes memory capacity and memory architecture a real constraint, not a side issue. Penguin’s pitch leans into that. It is not selling the main accelerator. It is selling part of the plumbing around it — the kind of thing enterprises buy when they want usable AI systems, not just raw chips. That makes PENG more of an enterprise AI deployment play than a pure semiconductor story. (ir.penguinsolutions.com) ### What is MRAM betting on? Everspin is a much narrower story. MRAM makes non-volatile memory — memory that keeps data without power — and it is pushing that into edge AI, industrial, aerospace, and defense systems. In March it launched its UNISYST MRAM family for embedded systems, explicitly targeting edge AI and mission-critical designs. Then in late April it reported preliminary Q1 2026 results and highlighted a subcontract that can total about $40 million. That is why the stock started moving — not because it suddenly became a core data-center winner, but because a tiny memory specialist landed a contract big enough to change the near-term narrative. (ir.penguinsolutions.com) ### Is MRAM really a data-center AI play? Not really — at least not in the same way people mean when they say “AI infrastructure.” Everspin’s products fit better in edge and embedded systems where persistence, power efficiency, and durability matter more than raw hyperscale volume. Think of it less like the main warehouse and more like rugged gear at the edge of the network. That can still matter a lot. But the catch is that investors chasing “AI memory” can blur together very different markets. (investor.everspin.com) ### Why is MXL in this basket? MaxLinear is the clearest networking-and-interconnect name of the three. In its April 23 Q1 2026 report, revenue hit $137.2 million, up 43% year over year, and infrastructure revenue jumped 136% on optical product traction. Management also raised full-year optical data-center revenue guidance to $150 million to $170 million. That is a much more direct “AI buildout needs faster links” thesis. If GPU clusters are the compute engine, MXL is part of the wiring that keeps those clusters fed. (investor.everspin.com) ### So why are traders grouping them together? Because the market wants second-order winners. Once the obvious GPU trade gets crowded, people look for suppliers one layer down. But this basket is messy. PENG is an enterprise AI systems and memory-architecture story. MRAM is an edge and defense memory story. MXL is an optical interconnect story. Same macro tailwind — very different customer bases, cycles, and failure modes. (investors.maxlinear.com) ### What should investors actually watch? Watch whether AI spending keeps broadening from chips into full-stack deployment. For PENG, that means enterprise and sovereign AI wins. For MRAM, it means turning niche design wins into repeatable revenue. For MXL, it means optical data-center ramps staying real and not just one-quarter enthusiasm. The bottom line is simple — these are not interchangeable AI picks. They are three different ways to bet that the bottlenecks around AI are moving outward from the GPU. (ir.penguinsolutions.com)