Product1 publisher3 min readPublished
Kepler claims 2nm-class SRAM density out of 28nm fabs without buying EUV
Kepler Computing has $468mn, up to $245mn committed by the US Commerce Department, and mini fabs running inside GlobalFoundries, which is more than most memory startups get and still short of anything a buyer can schedule against.
The Product Desk · Product desk
What happened
- Kepler Computing, founded in San Jose in 2018, came out of stealth after more than seven years, claiming an HBM and SRAM architecture that raises density without EUV lithography and runs in existing fabs.
- It has raised $468mn from backers including GlobalFoundries, Intel Capital, AMD Ventures, Baillie Gifford and Bill Gates's Gates Frontier fund, and in July the US Commerce Department committed up to $245mn.
- Kepler says its SRAM reaches the density of 2nm or 3nm chips with no EUV spend, and CEO Debo Olaosebikan says the HBM roadmap began after ChatGPT's 2022 launch and now runs in parallel with SRAM.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Kepler's own stated condition is production at scale, so this adds no wafers to next year's supply and gives nobody a reason to reprice a memory line for the builds they are quoting now.
- contradiction Kepler pitches relief for the memory shortage while its investor Intel Capital says it never assumed a DRAM or SRAM replacement; one framing sells to buyers, the other to fabs, and they imply different timelines.
- capability If the materials approach holds at volume, incremental memory supply could come out of fabs that already exist rather than out of new multibillion-dollar builds, which widens the list of plausible suppliers.
- exposure Commerce put up to $245mn of public money behind the approach in July, so scaling risk that would normally sit with equity backers is now shared with taxpayers.
The person who has to do something with this is holding a memory line item for next year's builds and deciding whether to move it. The evidence says leave it where it is, and Kepler's own condition is the reason: the shortage help arrives only if the company can make this at scale, according to Wired's account [3].
The $468mn raised privately plus the Commerce commitment of up to $245mn comes to as much as $713mn [15], with the federal money alone worth roughly 52 percent of everything the investors put in [16] -- enough to show that capital was never the constraint here. GlobalFoundries' $50mn is about a tenth of the private raise [17], but its more significant contribution is the line it opened up: much of the testing sits in GlobalFoundries' Singapore facility [8], and for two years Kepler has been running what it calls mini fabs there, producing memory chips in conjunction with GlobalFoundries' 28nm chips, with further test work in Burlington, Vermont [9].
That is where the claim gets genuinely testable. Kepler says stacking plus a proprietary material buys density without EUV [2], and specifically that its SRAM hits the density of 2nm or 3nm chips with no EUV investment [4]. Made on a 28nm-class line [9]. If that survives volume, the node number on a datasheet stops working as the shorthand for how much cache you get.
The coverage lays out the raise, the fab partner, the density claim and the architecture, but it stops short of a yield figure, a capacity number, a ship date, or a named memory customer [18]. Those four are the ones that turn a materials result into supply, and none of them are public.
A sorting rule for any claim of this shape: name which bottleneck it removes. Lithography tool access, wafer starts, advanced packaging, or your own qualification cycle. Kepler's public evidence speaks to the first and, through the 28nm mini fabs, hints at the second [9]. The record says nothing about the last two. A supplier who clears the bottleneck you do not have is a reasonable investment and an irrelevant purchase order.
Right now, this serves GlobalFoundries, whose CMOS chief Ed Kaste describes the approach as a new materials system with multigenerational scaling potential that needs neither new fab systems nor very expensive lithography equipment [12]. That is fab economics, and a good deal for a company that sells mature nodes. Kepler's own pitch is broader: that the accelerated computing market should not have to wait for new memory fabs, because existing ones can be made to yield more supply [19]. Meanwhile SK Hynix and Micron are the ones spending billions on the fabs that will set what is actually buyable [13], and Wired notes both that fab construction is expensive and slow and that memory demand runs in cycles [14].
A yield number attached to a named customer is the figure that would move a purchase order, and so far nobody has produced either one.
What to watch
- A yield or wafer-start figure attached to the Singapore mini fabs, which is what turns a density claim into a supply claim.
- A named memory customer or qualification win at an HBM buyer rather than at a fab partner or investor.
- How much of Commerce's up-to-$245mn actually flows to Kepler, and what it is paid against.