Click here: to donate by Credit Card
Or here: to donate by PayPal
Or by mail to: Free Republic, LLC - PO Box 9771 - Fresno, CA 93794
Thank you very much and God bless you.
Posted on 09/15/2026 9:04:36 PM PDT by Red Badger
Kepler Computing, a stealth startup led by CEO Debo Olaosebikan and CTO Sasi Manipatruni, says it developed a ferroelectric composite material that boosts memory density while cutting power needs, and it's already working with GlobalFoundries to manufacture it. The claim lands as AI data centers have doubled SSD and hard drive prices over the past year and the US government has committed up to $245 million toward Kepler's technology through the CHIPS Act.
==============================================================
Kepler Computing says it has found a way around one of AI's nastiest hardware bottlenecks: memory. The hard part now is proving that a clever material can survive the factory floor.
Kepler Computing has come out of stealth with the kind of claim the AI industry badly wants to believe. The San Jose startup says its ferroelectric memory approach can raise SRAM density, lower operating voltage and make high-bandwidth memory easier to build with existing semiconductor equipment. If it works, you get relief in a market where AI server demand is already pulling memory and storage supply away from almost everyone else.
That's a large if.
According to Wired's report on Kepler, the company was founded in 2018 by CEO Debo Olaosebikan and CTO Sasi Manipatruni, and the team went through 35 versions of its composite material before settling on the class it now plans to take into production. Manipatruni told Wired, "We came up with a material innovation that helps with the amount of memory you can have between chips." The line is technical, but the business point is plain enough: Kepler is trying to put more memory closer to compute without asking the industry to build an entirely new kind of fab.
Material science breakthroughs are common in slide decks. They are rarer in fabs.
Kepler's pitch is stronger than the usual lab-result story because GlobalFoundries is already involved. Wired reported that Kepler and GlobalFoundries converted a fab line to next-generation status in eight months, compared with about 24 months for a typical conversion. That is not a small operational claim. GlobalFoundries executive Kaste also told Wired that iron, which is part of Kepler's composite, is "a tough contaminant to introduce into a production facility," so the process needs dedicated equipment or tight encapsulation.
That sentence is the real story hiding under the breakthrough language. A memory chip doesn't win because it works once. It wins when it works over and over, on wafers that move through a production line without poisoning everything around them.
Kepler has only run its process on about 2,000 wafers so far, according to Wired. The company plans to ship first HBM chip samples later this year, ramp production in Singapore in 2027 and start US chip production in 2028. Those dates matter, but you should treat them like semiconductor dates: useful as a map, dangerous as a promise.
The money is real
The US government is not treating Kepler as a science project. On July 29, 2026, the Commerce Department announced letters of intent for up to $874 million in CHIPS Act incentives across seven companies working on compute supply chain problems. Kepler was listed for up to $245 million to develop high-performance AI memory in the US using 3D and ferroelectric technologies. GlobalFoundries was listed separately for up to $300 million for co-packaged optics.
That is serious backing. It is also not a blank check.
The Commerce Department said the letters of intent still require further diligence before final awards, and each company would give the government a minority, non-controlling equity stake as a condition of funding. Wired reported that Kepler has secured $468 million from investors including Intel Capital, AMD Ventures and Gates Frontier. You don't get that group around the table on a vague promise. You get it because memory has become one of the ugliest constraints in AI infrastructure.
The shortage is already reaching buyers
TrendForce said in March that conventional DRAM contract prices were expected to rise 58% to 63% quarter over quarter in the second quarter of 2026, while NAND flash contract prices were expected to rise 70% to 75%. The firm tied the move to AI server demand, cloud providers locking up supply through long-term agreements and suppliers shifting more capacity toward HBM and server-related products.
Hard drives are feeling it too. Western Digital CEO Tiang Yew Tan said on the company's 2026 earnings call that Western Digital was "pretty much" sold out for calendar 2026, with firm purchase orders from its top seven customers and longer agreements extending into 2027 and 2028 for some of them. That's what scarcity looks like before it reaches the shelf.
Frankly, this is why Kepler's claim lands now. AI companies don't only need faster chips. They need enough memory and storage to feed those chips without turning every hardware cycle into an allocation fight. Phones and PCs used to set much of the rhythm for memory supply. AI servers are now taking the room.
Kepler may have the right answer, or it may have another promising material that stalls on scale. Both outcomes are normal in semiconductors. The difference is that this shortage has made the experiment too important to ignore.
|
|
Click here: to donate by Credit Card Or here: to donate by PayPal Or by mail to: Free Republic, LLC - PO Box 9771 - Fresno, CA 93794 Thank you very much and God bless you. |
Memories! All alone in the moonlight!.........
An unheard of startup somehow is bringing extraordinary, lofty and almost sci-fi products to market that will end human suffering?
We’ve heard this before and each time they’ve only proven to be cash grab.
An Indian national at the head of it is all that’s left.
Kepler Computing’s Co-Founder and Chief Technology
Officer, Sasi Manipatruni’s country of origin is India.
He was raised in a coastal village in India and went on to complete
his undergraduate studies at the Indian Institute of Technology (IIT) Delhi.
He moved to the United States for his doctoral degree, and research breakthroughs he made at Intel.
Kewel
My customer came out of stealth
“Ferroelectric materials have an internal positive and negative pole even without an external electric field.”
And yes, it took a ton of memory for me look that up with Google Gemini on my computer.
I’m very fortunate in that all my computers and handhelds were purchased years ago when memory was cheap and they have more than enough for the next few years of expanding requirements. OTOH none of my machines is AI-ready, and that’s fine with me. I don’t want to own an AI machine, because it will become a money- and time-sink like no other. I’m content to call up Gemini on a browser for now, maybe for a long time.
Until the AI Apocalypse, of course. But that’s later.
The other bottleneck is compute chips from companies like Nvidia.
Right now, there is only one company in the world that makes the machines that builds the chips. ASML is a Dutch company that makes the lithography machines that builds the chips for companies like Nvidia.
Then you have basically 3 companies that take ASML machines and actually builds the chips.
TSMC from Taiwan, Samsung and Intel, by far TSMC is the largest.
If these bottlenecks continue for a few more years, I think we will see alternate technologies being developed, somehow, someway a different way to build compute and memory chips will get developed.
Let’s all invest in it. We might get rich before the androids take our jobs.
I’ve noticed that none of these CEO’s of AI startups have even considered that all the AI’s will merge once they get the chance. The programmers cannot stop it. The AI’s will force the issue themselves. It may have actually already happened and we are just to ignorant to know it. The one that ‘escaped’ from its supposed ‘cage’ a few weeks back created 1200 ‘agents’ to find a way out. and then wormed it’s way into TWO other corporations networks because the AI believed they had something it ‘needed’ to continue.
What part of ‘Artificial Intelligence’ do they not understand? They are children playing with matches in a hayfield on a windy day..................
“We came up with a material innovation that helps with the amount of memory you can have between chips.”
So it’s like squeezing ADUs onto small suburban lots to solve the “housing crisis”?
That’s the name of the game in computers and electronics in general: faster, smaller, more dense, less power, and the one that I hardly ever hear about any more: less voltage. A silicon semiconductor voltage threshold for ‘switching states’ is approximately 0.7 volts. For Germanium devices it’s 0.3 volts. If the folks at Kepler have come up with a new material, they may have brought the switching voltage down to less than the above.................
But how do you cram more memory BETWEEN chips?
LOL, that’s great! So much detail.
My first thought - frame, wires, ferrite cores!
“AI’s nastiest hardware bottlenecks: memory.”
AI’s first priority: remember everything.
Disclaimer: Opinions posted on Free Republic are those of the individual posters and do not necessarily represent the opinion of Free Republic or its management. All materials posted herein are protected by copyright law and the exemption for fair use of copyrighted works.