Google just fired a shot across the bow of the memory chip industry. And the companies most likely to feel the impact — Micron Technology and Samsung Electronics — may not have a good answer for it.
At its annual Cloud Next conference in early April, Google unveiled its seventh-generation Tensor Processing Unit, called Ironwood, alongside a striking disclosure: the company has begun designing its own custom high-bandwidth memory chips. The announcement, first analyzed in detail by Motley Fool, signals a potential structural shift in how hyperscale cloud providers source the components that power artificial intelligence workloads. If Google succeeds in building its own memory, the ripple effects could reshape the economics of an industry that has spent decades consolidating into a tight oligopoly.
The memory chip market is dominated by three players: Samsung, SK Hynix, and Micron. Together they control virtually all production of DRAM and NAND flash memory worldwide. High-bandwidth memory, or HBM — the specialized, vertically stacked DRAM used in AI accelerators — has become the hottest product in semiconductors over the past two years. SK Hynix currently leads in HBM production and has been the primary supplier to Nvidia. Micron has been aggressively ramping its own HBM output. Samsung, despite its massive scale, has struggled with yield issues and has been playing catch-up.
Now Google wants to cut all three out of the equation. Or at least reduce its dependence on them.
The logic isn’t hard to follow. Google’s TPUs are custom-designed chips built specifically for AI training and inference. The company has been designing its own silicon for over a decade, steadily reducing its reliance on merchant chip suppliers like Nvidia and Intel. Custom chips let Google optimize performance per watt and performance per dollar in ways that off-the-shelf hardware can’t match. Memory is the next frontier. By designing its own HBM, Google could tightly integrate memory and compute in ways that improve data throughput, reduce latency, and lower the total cost of ownership for its sprawling data center operations.
That’s the theory. The practice is far more complicated.
Designing HBM is one thing. Manufacturing it is another entirely. High-bandwidth memory requires advanced packaging techniques, including through-silicon vias and micro-bump bonding, that are extraordinarily difficult to execute at scale. Even Samsung — a company with decades of memory fabrication experience and some of the most advanced fabs on the planet — has stumbled on HBM yields. Google doesn’t own fabs. It would need to partner with a contract manufacturer, most likely TSMC or Samsung’s foundry division, to actually produce custom memory chips. The irony of potentially contracting Samsung to manufacture memory chips designed to displace Samsung’s own products is not lost on industry observers.
But Google has pulled off this kind of vertical integration before. Its Tensor Processing Units are fabricated by Broadcom and manufactured at TSMC. The company designs the architecture, owns the intellectual property, and outsources the physical production. There’s no reason it couldn’t apply the same model to memory. The question is whether the performance and cost advantages of custom HBM justify the enormous R&D investment required.
For Micron, the timing is particularly painful. The Boise, Idaho-based company has staked its near-term growth strategy on HBM. CEO Sanjay Mehrotra has repeatedly told investors that HBM represents a massive revenue opportunity, with the company targeting billions of dollars in HBM sales over the next several years. Micron’s stock has surged on AI enthusiasm, with investors betting that insatiable demand for AI training chips would translate into equally insatiable demand for the memory those chips require. Google’s announcement complicates that narrative. If the largest cloud companies start designing their own memory, Micron’s addressable market shrinks — potentially significantly.
Samsung faces a different but equally uncomfortable problem. The South Korean conglomerate is already losing HBM market share to SK Hynix. If Google — one of the world’s largest buyers of data center hardware — begins sourcing custom memory, Samsung loses a major customer at precisely the moment it can least afford to. Samsung’s semiconductor division has been under intense pressure, with operating profits declining and internal restructuring efforts underway. Another blow to its memory business would compound an already difficult situation.
SK Hynix, interestingly, may be the least affected in the short term. Its dominant position in HBM supply to Nvidia gives it a degree of insulation. But even SK Hynix should be watching carefully. If Google’s custom memory initiative succeeds, Microsoft, Amazon, and Meta won’t be far behind. All four hyperscalers have been on a multi-year push to design custom silicon. Amazon has its Graviton and Trainium chips. Microsoft is developing Maia for AI workloads. Meta has its own custom inference accelerators. Custom memory is a logical next step for all of them.
The financial implications are substantial. The HBM market is projected to exceed $25 billion in annual revenue within the next few years, according to industry estimates. If even a fraction of that demand shifts from merchant suppliers to custom-designed chips, the impact on Micron’s and Samsung’s top lines could be material. Memory is a cyclical business with razor-thin margins during downturns. Losing volume at the top of the cycle — when prices are high and margins are fat — would be especially damaging.
There’s a broader strategic dimension here too. For years, the semiconductor industry’s value chain has been relatively stable: chipmakers design and manufacture components, and systems companies buy them. The rise of custom silicon at hyperscale cloud companies has been steadily eroding that model. Google, Amazon, and Microsoft now design their own CPUs, AI accelerators, and networking chips. They’re becoming semiconductor companies in all but name. Custom memory represents the next layer of the stack to be internalized.
Not everyone is convinced this will play out as dramatically as the bears fear. Memory manufacturing is among the most capital-intensive activities in the technology industry. Building a single advanced DRAM fab costs upward of $15 billion. Even Google, with its vast resources, may find that the economics of custom memory don’t pencil out at scale — especially when existing suppliers are investing tens of billions of dollars annually to expand capacity and improve performance. There’s also the question of whether custom memory delivers enough of a performance advantage over standard HBM to justify the investment. Google’s TPUs are used primarily within Google’s own cloud infrastructure. The volumes, while large in absolute terms, are small relative to the total memory market. Merchant suppliers could simply absorb the loss and move on.
But that argument assumes Google remains the only hyperscaler pursuing custom memory. If Amazon, Microsoft, and Meta follow suit — and history suggests they will — the cumulative impact on merchant memory suppliers could be severe. The playbook is familiar. When hyperscalers started designing custom CPUs, Intel’s data center business cratered. When they started designing custom AI accelerators, Nvidia’s dominance came under question (though Nvidia has so far maintained its lead through sheer performance advantages). Custom memory could follow the same trajectory: slow adoption at first, then a rapid shift as the technology matures and costs come down.
Wall Street has been slow to price in this risk. Micron’s stock still trades at a premium to its historical valuation, buoyed by AI optimism. Analysts have largely treated the HBM opportunity as a given, with few questioning whether demand from hyperscalers will remain captive to merchant suppliers. Google’s announcement should prompt a reassessment. The risk isn’t that custom memory replaces HBM overnight. It’s that the threat of custom memory gives hyperscalers bargaining power they didn’t previously have — power to demand lower prices, better terms, and priority allocation from suppliers who can no longer take their business for granted.
That dynamic alone could compress margins across the memory industry, even if custom memory never achieves significant market share.
Google, for its part, has been characteristically tight-lipped about the details. The company hasn’t disclosed timelines for when custom memory might enter production, nor has it identified manufacturing partners. What it has made clear is that Ironwood, its latest TPU, represents a significant leap in AI compute density — and that memory bandwidth is the key bottleneck it’s trying to solve. Custom memory is the most direct path to solving it.
The semiconductor industry has been here before. Every few years, a major customer decides it can do better than its suppliers. Sometimes it works — Apple’s custom chips for iPhones and Macs have been spectacularly successful. Sometimes it doesn’t — Amazon’s Graviton CPUs are competitive but haven’t displaced x86 processors in most workloads. Memory is a harder problem than logic chips, with physics constraints that limit how much customization is possible. But Google has the engineering talent, the financial resources, and the strategic motivation to try.
Micron and Samsung should be worried. Not because custom memory will kill their businesses tomorrow. But because the ground beneath them is shifting, and the companies that buy the most memory in the world are no longer content to simply write checks. They want control. And in the semiconductor industry, control eventually flows to whoever designs the silicon.
WebProNews is an iEntry Publication