> daily_signal(2026_06_29)

Samsung and SK Hynix put $590 billion behind the AI memory shortage, and warned the relief you're waiting on won't arrive before 2028.

PickBits Daily Signal · Monday, June 29, 2026

By Mark Pickering · 7 min read · June 29, 2026

// tl;dr

Samsung and SK Hynix unveiled a combined ~$590 billion plan to expand memory production, but said relief won't come before 2028. The package is roughly 800 trillion won for four new fabs in southwest South Korea, 81 trillion won for a packaging center, and 30 trillion won over 15 years for next-gen chips, backed by President Lee Jae Myung. The two firms make ~80% of the world's high-bandwidth memory, and say demand outruns supply for years; prices are projected to rise another 40–50% this quarter.
The FTC cleared Elon Musk, named personally rather than SpaceX, to acquire Mesh Optical, a data-center optical-networking startup. Founded by former SpaceX engineers, its Alpha C1 transceiver hits 1.6T and 800G speeds at about a third the power of rival modules. It puts the interconnect layer of the AI stack — after memory and power — under a single owner.
California is set to let unspent CalSHAPE money, the funds for cleaning the air in public schools, revert to utilities rather than upgrade school ventilation. The California Schools Healthy Air, Plumbing, and Efficiency program is run by the California Energy Commission; amid the 2026–27 budget squeeze, money districts don't claim goes back to utilities. Schools can check energy.ca.gov for unclaimed funds.
AI for good: researchers in New Zealand are training AI on discarded bowel-screening samples to flag bowel-cancer risk at up to ~90% accuracy, no colonoscopy needed. The national programme's FIT stool tests are normally thrown away after checking; the AI reads the gut-microbiome bacteria in them, modeling multi-species patterns simpler tests miss. A separate microbiome stool-test line reports catching ~90% of colorectal cancers, and AI-assisted colonoscopy (already deployed over a year at sites like Mater) catches hard-to-see polyps in real time.

The first two stories come back to the same question: who controls the physical hardware AI runs on. We've spent weeks tracking the memory shortage as it crept into consumer prices, Apple and Microsoft raising them, the new MacBook losing its aggressive price, a camera maker priced out of its own components. Now the supply side answers with the largest number yet, $590 billion, and the same firms tell you not to expect relief for two more years. The FTC cleared Musk to pull data-center optics in-house, after memory and power already consolidated. Then the day turns local: a few million dollars for the air in California classrooms that reverts to utilities if nobody claims it. And it ends on the other half of the AI story, the part that isn't about ownership at all: AI trained on stool-screening samples that would otherwise be thrown away, learning to flag bowel cancer earlier from a test people already take.

Today: Samsung and SK Hynix committed $590 billion to memory but ruled out relief before 2028, the FTC cleared Elon Musk personally to buy data-center optics startup Mesh Optical, California is poised to send unspent school clean-air funds back to utilities, and researchers are training AI on discarded bowel-screening samples to flag cancer risk at up to ~90% accuracy without a colonoscopy.

1. Samsung and SK Hynix put $590 billion behind the memory shortage, and said relief won't come before 2028.

The supply side finally answers the AI memory crunch, with a number and a timeline.

Samsung and SK Hynix unveiled a combined plan of roughly $590 billion to expand memory-chip production: about 800 trillion won for four new fabs in southwest South Korea, 81 trillion won for a packaging center, and 30 trillion won over 15 years for next-generation chips, backed by President Lee Jae Myung's regional-growth push. The two companies make roughly 80% of the world's high-bandwidth memory, the kind AI training depends on. The plan is a direct response to demand that has outstripped what they can make, and the honest part is the timeline they attached to it: demand outruns supply for years, and relief is not expected before 2028. Memory prices are projected to climb another 40–50% this quarter alone.

This is the supply-side answer to the shortage we've been tracking from the demand side for weeks, the one that has already pushed up the price of laptops and phones and squeezed companies whose products depend on the same chips. It's a huge commitment, and it comes with a built-in catch. Fabs take years to build, so the money and the relief run on different clocks: the spending starts now, the new supply shows up at the end of the decade. So don't fixate on the headline number, look at the gap between those two dates. For the next two years, anyone whose costs ride on memory is buying into a market the people who make it have told you stays tight.

Why this matters: If anything you build or buy rides on memory, a product, a fleet of machines, a cloud bill, the two companies that make most of it just told you the squeeze lasts into 2028. Waiting for prices to settle is the wrong move here, because the people with the clearest view of supply just told you they won't, not soon. Action this week: Re-price any 2026–27 plan that quietly assumed flat memory costs, using the projected 40–50% jump this quarter as your baseline rather than last year's number. If you buy memory or high-bandwidth memory directly, get a supply-and-pricing review on your vendor's calendar this quarter while you still have lead time, and lock what you can. And if you've been deferring a hardware refresh on the theory that prices dip, treat that as a bet against the only two firms who actually know.

the-decoder.com: Samsung and SK Hynix plan $590 billion chip investment as AI demand sends memory prices soaring (June 2026)

2. The FTC cleared Elon Musk, personally, to buy the optics startup that wires his AI data centers.

After memory and power, the interconnect layer gets pulled under a single owner.

The FTC cleared Elon Musk, named personally in the approval rather than SpaceX or his AI company, to acquire Mesh Optical Technologies, a data-center optical-networking startup founded by former SpaceX engineers. Its Alpha C1 transceiver reaches 1.6T and 800G speeds at about a third the power of rival modules, which is exactly the kind of part the biggest AI systems are short on. Optical interconnect — the high-speed links that move data between chips inside a data center — is becoming a bottleneck the same way memory and power did, as the industry races to replace copper cabling with glass fiber to cut energy use.

The story isn't one acquisition; it's the pattern. The layers of the AI hardware stack, the chips, the memory, the power, and now the wiring between them, keep getting absorbed by the same small set of people building the largest systems. Sure, there's an upside: one owner and one roadmap is how you make a stack this complicated actually run. But I keep landing on what it crowds out. When one person owns the wiring and the memory and the power, there's no neutral ground left for anyone else to build on. "We own the whole pipeline" sounds great if you're the one who owns it, and a lot worse if you're trying to compete.

Why this matters: If you build on AI infrastructure, the layer you assume is a neutral, competitive market, the wiring, the same way memory was, is quietly becoming something one owner controls. A part that runs at top speed on a third of the power isn't a niche networking detail; in a power-constrained data center it's leverage, and Musk just took it off the open market. Action this week: Map which layers of your own AI supply chain — interconnect, memory, power — are consolidating under a single owner, and write down where you have a real second source versus where you only think you do. If you procure data-center networking, ask your vendors for power-per-bit numbers on their optical parts before your next buildout, because that figure, not raw speed, is where the squeeze shows up.

datacenterdynamics.com: Elon Musk gets FTC nod to acquire data center optics business Mesh Optical (June 2026)

3. California is about to let school clean-air money revert to utilities.

A public-health choice hiding in a budget default: claim it, or it goes back to the power companies.

Amid California's 2026–27 budget squeeze, unspent money from CalSHAPE, the California Schools Healthy Air, Plumbing, and Efficiency program, run by the California Energy Commission, is set to return to utilities rather than fund ventilation and air-quality upgrades in public schools. The mechanism is a quiet default: money a district doesn't claim and spend doesn't roll forward and doesn't stay earmarked for schools; it reverts to the utilities that funded the program. The practical effect is that the air in a lot of classrooms stays as-is, and the money that was set aside to improve it goes somewhere else.

The dollar figure is small next to the other two, but the way the cut happens is what got my attention. Nobody has to vote to take clean air away from kids. It happens on its own, through a deadline, unless someone in each district actively claims what's already been allocated. That's the kind of choice that never makes news, because there's no announcement, just funds that were available in June and gone by the wind-down. It is also, unlike most of what we cover, genuinely fixable by one person filing the right paperwork.

Why this matters: If you have any connection to a California public school, whether you work at one, your kid attends one, or you sit on a board, there is money allocated for the air in those buildings that is about to go back to utilities unless someone claims it. This is the rare civic story where the action is small, concrete, and actually changes the outcome. Action this week: If you work at or with a California public school or district, go to energy.ca.gov and check whether your site has unclaimed CalSHAPE ventilation or plumbing funds, and start the application before the wind-down deadline. If you don't run a district but know someone who does, send them this with one sentence: check whether your CalSHAPE money is claimed before it reverts. The deadline, not the budget, is what costs schools this money.

canarymedia.com: California schools or utilities: who gets the unspent clean-air money? (June 2026)

4. AI for good: training cancer detection on the stool tests we already throw away.

The other half of the AI story: not who owns the stack, but what it quietly does for people.

Researchers in New Zealand are training AI on a sample most people never think about twice. The national bowel-screening programme mails out a simple FIT stool test; once the lab reads it, the sample is normally discarded. The team is training AI on those leftover samples to analyze the gut microbiome, the bacteria in the sample, modeling complex multi-species patterns that simpler analysis can't see. Early results flag a person's bowel-cancer risk at up to ~90% accuracy, with no extra test and no colonoscopy. It is part of a broader shift: a separate microbiome stool-test approach reports detecting around 90% of colorectal cancers, and AI that watches a colonoscopy in real time is already deployed (over a year at sites like Mater) to catch hard-to-see polyps.

This is the counterweight to the rest of the slate. The same class of technology that the memory, optics, and data-center stories are all fighting to own is, in a lab in New Zealand, doing something unambiguously useful: pulling far more signal out of a test millions of people already take, so cancer gets caught earlier and less invasively. It is early-stage research, not a product you can ask for yet, but the direction is the point, and the action it asks of you is the easiest on this whole list.

Why this matters: Bowel cancer is one of the most survivable cancers when it's caught early and one of the deadliest when it isn't, and the single biggest barrier to early detection is that people skip screening, often because they dread a colonoscopy. The promise here is more early-warning signal from a non-invasive test people are already mailed, which is exactly the kind of friction-free screening that actually gets done. Action this week: If you're in the age band that receives a bowel-screening kit, do the next one instead of letting it sit on the counter, that ordinary test is the raw material this AI turns into an early risk flag. If you run a screening programme or lab, ask whether discarded FIT samples can be retained with consent for microbiome-AI research. And send this to the person in your life who keeps putting their screening off.

rnz.co.nz: AI promises breakthrough in bowel cancer detection (June 2026)
sciencedaily.com: New stool test detects 90% of colorectal cancers (2026)
mater.org.au: Artificial intelligence improves bowel cancer detection (May 2026)

» What to watch this week

Whether memory prices actually post the projected 40–50% Q3 jump, and how fast it shows up in device pricing. The tell is whether more hardware makers follow Apple and Microsoft in raising prices, or start quietly cutting specs to hold a price point.
Whether the FTC's personal-name framing on the Mesh Optical clearance becomes a pattern for Musk's other acquisitions. An approval that names the individual rather than the company is unusual, and worth watching as a signal of how his AI hardware stack gets assembled.
Whether any California district publicizes its unclaimed CalSHAPE balance before the wind-down. A single district making noise about reverting funds is how a quiet default becomes a fight worth having.
Whether the bowel-screening AI holds its ~90% accuracy when it's validated beyond the initial sample set, and whether a screening programme moves to retain FIT samples for it. The tell on whether this becomes real clinical screening is independent validation plus a programme willing to consent and bank the samples the model needs.

Tomorrow's signal lands here.