Samsung's patent crown, a pay dispute, and the memory crunch nobody planned for
Samsung has held the US patent crown for four straight years. Inside its fabs, a US$402,000 bonus has reopened a fight over who built the AI era, and a separate warning says the world will run short of memory by 2030.
On 15 July 2026, the year's mid-point list of US patent grants confirmed what the hardware supply chain has suspected for some time: Samsung Electronics has now held the top spot for four consecutive years, overtaking IBM's decade-long run in 2022 and refusing to give it back through 2023, 2024 and 2025. The Korean conglomerate is the only non-US company to top the list in the modern era, and it did so while IBM's own filings dropped year over year.
The patent headline, though, is the polite version of the story. The harder version is being fought out in Suwon and Hwaseong, where a US$402,000 retention bonus offered to chip-making staff has produced an open feud with the device division that makes the Galaxy phones and the televisions. The same week, a fresh supply forecast warned that by 2030, demand for DRAM could exceed supply by 28.7 exabytes, roughly half of current annual global production. Three different facts. One underlying object: the silicon whose absence would quietly stall the AI build-out, and the labour force that makes it.
The patent scoreboard, and what it does not measure
The list is straightforward on its face. Samsung has been the largest recipient of US patents granted in each of 2022, 2023, 2024 and 2025. IBM held the title every year from 2012 through 2021. Samsung's streak is the first sustained non-US run in the modern record and arrives at a moment when the company's R&D is tilted heavily toward memory, advanced packaging and the high-bandwidth memory (HBM) that Nvidia's accelerators now require in volume.
The number itself, however, tells less than it appears to. US patent counts reward filings, not commercialised inventions. A patent granted is a legal instrument, not a product on a wafer. Samsung's lead reflects the breadth of its portfolio across chips, displays and consumer electronics, and the cost it is willing to absorb to keep that portfolio thick. It does not, on its own, tell a reader whether Samsung has closed the gap with TSMC in leading-edge logic, or with SK hynix in HBM3E and HBM4 yield. The interesting question is what the company is doing with the headroom the patents buy it.
A bonus, and a fissure
The bonus fight, reported by the South China Morning Post on 15 July 2026, is more concrete. Samsung offered roughly US$402,000 to engineers in its chip division as a retention package tied to the AI cycle. Device division staff, who build the products most consumers actually buy, were not included on the same terms. Within days, the disparity had produced an internal row serious enough to surface in Korean media: chip engineers accusing device colleagues of free-riding on AI-era margins, device engineers accusing the chip side of capturing rents that belong to the company as a whole.
The argument is not really about the money, although the money is large. It is about who in a vertically integrated conglomerate gets to claim credit for a cycle that is, at bottom, an infrastructure cycle. AI revenue flows through Nvidia's accelerators, but the accelerators are useless without HBM, and the HBM market is split between SK hynix, Micron and Samsung. Samsung's device business is what pays for the corporate overhead, the marketing budget and the brand that lets the chip sales force walk into a hyperscaler procurement office. Strip the device division out of the story and the chip division is a memory supplier in a duopoly with a Korean neighbour; keep it in and Samsung is one of the few companies in the world that can credibly bid for the whole stack.
Read either way, the dispute is a stress test of the conglomerate model that built Korean industrial policy. The chaebol structure, with cross-subsidies between divisions and a holding company at the centre, has always depended on a quiet consensus that no unit is more important than the firm. A US$402,000 wedge breaks that consensus quickly.
The memory crunch that the bonus is trying to preempt
The third data point is the one with the longest fuse. Researchers now estimate that by 2030, DRAM demand could exceed supply by 28.7 exabytes, roughly half of current annual global production. The single biggest driver is AI. Training runs, inference fleets and the on-device caches that let a phone or laptop run a small model locally all consume memory at rates the industry's prior forecasts did not anticipate. The shortage, if it arrives as projected, will not look like the 2021 to 2022 automotive chip crunch. There is no easy substitution path. You can re-spin a microcontroller on a different process node; you cannot synthesise additional DRAM wafer starts inside a calendar quarter.
That is the structural backdrop against which the Samsung bonus exists. The company is not paying chip engineers more out of generosity. It is paying them to stay because every competitor in the memory market is trying to do the same thing, and because the supply curve has begun to bend against the buyers. Samsung's market share in DRAM has lagged SK hynix in recent quarters, particularly in HBM, where SK hynix's early lead with Nvidia set a template the rest of the industry is still trying to match. A retention package is the cheapest defensive move available while capacity ramps come online in 2027 and 2028.
What the three facts point to
Read together, the patent lead, the internal bonus feud and the projected 2030 shortage describe a single industry at the moment its input becomes its bottleneck. For two decades, memory was a cyclical commodity with fat margins in upyears and bankruptcies in downyears. AI has changed that. Memory is now a strategic input on roughly the same tier as advanced logic and packaging capacity, and the firms that control it enjoy a position the old commodity model did not allow.
The Korean state's industrial policy has, for thirty years, treated memory as the asset to defend and the asset to subsidise. The patent list suggests that posture is still producing output. The bonus dispute suggests the people inside the firm have noticed that the asset is worth more than the conglomerate's internal pay scales assumed. The 2030 supply forecast suggests the rest of the world is about to notice too.
What remains genuinely uncertain is the substitution question. HBM is currently built on DRAM die stacked inside a packaging flow that looks more like an advanced logic process than a memory process. Whether other architectures, stacked SRAM on the accelerator die, novel non-volatile memories, near-memory compute fabrics, can absorb part of the projected 28.7 exabyte gap before 2030 is open. The sources do not specify. What they do show is that Samsung, for the moment, is the firm with the most US patents, the most visible internal labour dispute, and the most to lose if the supply curve bends as the forecasts say it will.
This article sits at the intersection of three Monexus beats: industrial policy in East Asia, the hardware layer beneath the AI build-out, and the labour politics of a conglomerate whose divisions no longer agree on who pays for whom. The wire framing on the same day treated the three stories as separate items. We treat them as one.
Wire provenance
This editorial synthesis draws on the following public wire/social posts:
- https://x.com/xstats_feed/status/
- https://x.com/pirat_nation/status/
- https://t.me/SCMPNews
- https://en.wikipedia.org/wiki/Samsung_Electronics
- https://en.wikipedia.org/wiki/High_Bandwidth_Memory