Dear Aventine Readers, 

Lots of tech products are getting more expensive. In many cases the massive build-out of AI data centers is to blame. This week we look at what's behind that dynamic and what's needed to fix it. 

Also in the issue: 

Thanks for reading, 

Danielle Mattoon 
Executive Director, Aventine

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

The AI boom is extraordinarily expensive for tech companies. It’s now making things expensive for everyone else, too.

Prices for computer memory have nearly doubled in the last seven months, with the cost of some types of memory more than quadrupling. Prices for other computer components have risen too: The cost of solid-state drives has doubled; hard-disk drives are up by 50 percent; and processor prices have climbed about 15 percent. But it is the cost of memory — a component found in almost all electronic devices, from phones and computers to internet routers and cars — that is climbing fastest, creating a chain reaction of price increases in everyday tech products. 

Why is this happening? In short, the construction of AI data centers is consuming vast quantities of memory chips, which are used to hold the enormous amount of data that processors manipulate. Memory makers have been diverting manufacturing capacity to meet that demand, in turn reducing supplies available for consumer hardware and pushing prices up. “The rising cost situation is unprecedented,” said Yuanqing Yang, chairman and CEO of computer maker Lenovo, in the company’s most recent earnings call. “And [it’s] still not finished.”

Aventine spoke with industry experts to understand what is driving these price rises, how hardware manufacturers are responding and how the situation might shake out over the coming months and years.

The great AI spend

Late last summer, the big story about AI was whether the massive investment it was attracting was creating a bubble. Money was pouring into infrastructure to facilitate a technology that, sceptics argued, couldn't live up to the hype. Now the question is whether AI companies will be able to keep up with customer demand. 

Anthropic’s annual revenue run rate, for example, has increased from $9 billion at the end of 2025 to $30 billion in April. It now predicts its revenues could grow by 80 times just this year. “I hope that 80-times growth doesn’t continue because that’s just crazy and it’s too hard to handle,” said Dario Amodei, the company’s CEO, at a conference in early May. “I’m hoping for some more normal numbers.” 

The not-normal numbers suggest that the AI hype could be justified after all, sending tech companies rushing to increase investments in AI infrastructure. On first-quarter earnings calls over the past few weeks, Google, Amazon, Microsoft and Meta said they collectively plan to spend $725 billion in capital expenditure this year — a 77 percent increase on last year's record $410 billion. According to industry analyst SemiAnalysis, around 30 percent of that will be spent on memory. And that likely won’t be enough. Demis Hassabis, CEO of Google DeepMind, has called the lack of memory a "choke point" for the industry, arguing that demand for AI vastly exceeds what existing physical infrastructure can supply. 

The upshot is that these companies are "willing to pay a significantly higher premium" for the memory required to build out their projects, said Jeff Janukowicz, research vice president at IDC. And that drives up prices for everyone else. 

A zero-sum game

Not all memory is the same. Consumer devices use so-called dynamic random-access memory, or DRAM — the workhorse version of the hardware. Data centers, meanwhile, use high-bandwidth memory, or HBM, which stacks DRAM chips on top of one another to multiply the amount of data the memory can move at once.

Most memory fabrication plants, or fabs, can produce either DRAM or HBM from the same wafers — the sheets of semiconductor material from which chips are cut. But they must choose which to prioritize. Since the companies building AI data centers are willing to pay top dollar, memory makers are pivoting production from DRAM to HBM, said Janukowicz, to "focus on the places that will pay the most." HBM supply is also constrained by limits on advanced packaging — the processes of assembling the chips — though AI buyers are taking everything available regardless of price.

Fabs can produce only a finite quantity of wafers. And each HBM chip uses roughly three to four times as much wafer as a standard DRAM chip, meaning every HBM chip allocated to AI servers sharply reduces chips available for consumer and enterprise products. "There's only a limited amount of memory production capacity in the short run," said Chris Miller, author of the 2022 book “Chip War: The Fight for the World's Most Critical Technology” and an economic historian. "So if more chips go to data centers, there's fewer left for other applications."

The consumer electronics sector is experiencing the most immediate impact. "It's very difficult to absorb those price increases and still sell a $399 PC when the cost of memory may have jumped up $50 versus where it was just a couple of quarters ago," said Janukowicz. Rising memory prices could push mainstream laptop prices up by as much as 30 percent, according to TrendForce, a market intelligence firm. As a result, IDC has warned that the PC market could shrink by more than 11 percent in 2026 and the smartphone market by more than 12 percent.

The lower end of the market will be hit hardest, since memory makes up a larger share of costs in cheaper devices, Janukowicz said. Some manufacturers may also lower specifications, creating products with less memory, slower processors or worse screens to maintain a more affordable price point. Other consumer products — printers, Wi-Fi routers, smart TVs, robotic vacuums — all use memory and are likely to see price increases too. Industries with higher-margin products, like automotive and medical devices, will absorb the increases more easily because memory takes up a smaller share of their overall costs. 

The AI sector itself is also paying more. Microsoft's CFO, Amy Hood, attributed $25 billion of the company's 2026 capex increase to rising memory and component costs, while Meta cited "higher component pricing this year, particularly memory" for a $10 billion bump in its capex plans.

How to fix the memory shortage 

If everything else holds steady, the only way to create more memory is to build it — a sore subject for memory company bosses. The industry has built memory capacity before only to find itself with too much of it when demand softened. This happened most recently in 2023 after device makers stockpiled components during the Covid-19 pandemic only to face weaker consumer demand when things normalized. "While we saw this wave of AI coming last year, the industry still was kind of cautious in terms of really building new fabs," said Janukowicz. The question they’ve had to grapple with is whether AI demand is durable enough to justify a permanent increase in manufacturing capacity.

For now, the answer is a cautious yes. The three dominant memory makers — America's Micron, alongside South Korea's SK Hynix and Samsung, which together control around 95 percent of the global market — are all currently building new plants in the US and Asia. But new fabrication facilities take two to three years to build, and the new capacity isn't expected to come online until 2027 or 2028 at the earliest. The current supply situation, said Janukowicz, will "last out into 2027" as a result. Hyperscale customers are now seeking long-term agreements for memory, though, which may give the big three companies extra confidence to build still more capacity.

The lag also opens a window for Chinese firms to challenge the oligopoly. ChangXin Memory Technologies, founded in 2016 and based in Hefei, already controls about 5 percent of the global DRAM market. "Right now Chinese firms make up a small share of global memory production, and for DRAM they can't produce the most advanced chips," said Miller. "However, China is investing heavily and the global shortage creates an opening for them to begin selling to international markets." US export controls on advanced manufacturing equipment will continue to limit Chinese companies’ ability to produce cutting-edge HBM, but the shortage creates an opportunity in mainstream DRAM.

The other route to lower prices would be more abrupt: the AI bubble bursting. "If AI data center construction slows, memory prices will fall," said Miller. But for now, there is no sign of that. AI demand is soaring, company revenues are exploding and tech executives can't see an end to the buildout. "I think for the next 10 years, there will always be more demand than supply," said Google Cloud CEO Thomas Kurian. 

Some analysts argue that the demand for memory isn't a short-lived shock at all but a structural shift — one in which memory prices never return to pre-2025 levels. “The large memory players will invest new capital, they will put new supply in place,” said Janukowicz. “But it will be ... a very different environment moving forward.”

Drug development is headed to space. Varda Space Industries, a startup aiming to manufacture pharmaceuticals in orbit and return them to Earth, has struck its first commercial partnership. The company will work with United Therapeutics, a biotech company based in Silver Spring, Maryland, to investigate whether drugs can be manufactured differently in microgravity. As MIT Technology Review reports, the theory is that weightlessness changes how molecules crystallize, potentially allowing for the creation of drugs with better stability or other improvements. That could mean that therapies we already use could perhaps have longer shelf lives or be delivered in new ways — as a simple injection rather than through an IV drip, say. Varda has flown six test missions since 2024, conducting a mix of pharmaceutical and defense-related experiments. Working with United Therapeutics, it plans to launch small manufacturing capsules, as it calls them, aboard SpaceX rockets. The capsules will then autonomously process chemicals in orbit before being returned to Earth by parachute. Manufacturing drugs in space sounds extravagantly expensive. But some pharmaceuticals are so valuable that the economics may not be totally absurd. A kilogram of semaglutide, for example, can be worth more than $100 million at retail prices. And there is some scientific basis for thinking space could be conducive to drug manufacturing: Experiments aboard the International Space Station have shown that crystals can form differently in microgravity. Still, no company has ever manufactured products in space and sold them on Earth. So for now, count Varda’s efforts as purely proof of concept. 

An experiment in putting AI where there’s power. As an alternative to building gigantic data centers and then scrambling to supply them with electricity, Nvidia is exploring a different approach: building smaller data centers close to already existing electricity supplies. According to IEEE Spectrum, the company plans to build a network of 25 compact data centers, each consuming between 5 and 20 megawatts of power (about the same as a small town uses) near power substations. Individually, they are far smaller than the hyperscale AI facilities now being developed, many of which demand 100 megawatts or more. But the size allows them to take advantage of power that already exists. Electrical grids carry more power than is needed day-to-day because they are designed to handle rare moments of peak demand. This means a lot of capacity is unused. Nvidia’s data center network will capitalize on this by running a single, coordinated system using the excess power created by the 25 substations. If one or more substations experience high demand from other loads, the corresponding data center will shut down and other data centers in its network will compensate. It should be noted that this approach will only work to power inference — that is, running AI models; training the models requires more, and more predictable, power. Construction of the pilot fleet is expected to begin by the end of 2026.

Ford is tearing up its rulebook to build EVs affordably. After years of struggling to make money on electric vehicles, Ford is trying to fundamentally rethink how it designs and builds them. Its early EV push — including the heavily promoted electric Ford F-150 Lightning — struggled to achieve the mass-market appeal needed to compete with lower-cost Chinese rivals (if you are buying outside the US) or with Tesla. According to The Wall Street Journal, the company’s answer has been a secret internal project underway since 2022. Operating largely out of California and staffed by former Tesla and Apple engineers, the team was tasked with designing a new $30,000 electric pickup from scratch. Doing that has apparently required challenging almost every long-held assumption inside Ford. Along the way, the team has proposed dramatically reducing wiring, cutting hundreds of components, redesigning vehicle bodies to simplify manufacturing and adopting more iterative engineering methods inspired by Silicon Valley. That has reportedly created tension with longtime Ford executives and engineers accustomed to slower, more rigid development processes. But the new truck is expected to launch in 2027, with factory-built prototypes expected later this year from Ford’s Louisville plant. With EV incentives disappearing in the US and consumer appetite for EVs softening, the new vehicle will be a test of whether a century-old giant can provide the spark to reinvigorate an industry. 

I Work in Hollywood. Everyone Who Used to Make TV Is Now Secretly Training AI, from Wired
5,000 words, or about 20 minutes

Don’t be fooled by the headline: This isn’t about Hollywood, but about the emerging labor system underneath the AI boom. The author, Ruth Fowler, is a television writer and showrunner with a degree from the University of Cambridge who turned to AI gig work after struggling to maintain a stable income in the TV industry. Like thousands of others, she signed up for platforms such as Mercor, Outlier, Turing and Task-ify, which recruit humans to generate and evaluate the data used to train AI systems. The pitch is compelling: flexible hours, engaging work, pay that hits $150 an hour. But Fowler’s account suggests that the reality doesn’t match up. Her description of the industry is scathing, darkly funny and reminiscent of stories we’ve heard in the past about Uber and DoorDash. Workers endure interviews conducted by AI chatbots that themselves feel like training exercises. Young managers send motivational messages in the middle of the night. Highly skilled professionals compete frantically for a small number of tasks with ever decreasing pay. Some workers pull all-nighters simply to earn enough before the work disappears. And there’s something particularly bleak about experienced professionals in their 30s and 40s — writers, lawyers, academics — training AI systems that may automate large parts of their own industries. The piece is one-sided, but it captures the emotional experience of a new era of gig work.

Grate expectations: the troubled quest for tasty vegan cheese, from 1843
3,700 words, or about 15 minutes

Ever tried vegan cheese and decided you never want to eat it again? You’re not alone: It accounts for less than 1 percent of the US cheese market, a share that’s declining. The problem is that real cheese depends heavily on casein, a family of milk proteins responsible for the melt, stretch, texture and flavor we’ve come to expect. Replicating that using plant ingredients is extraordinarily difficult, so most vegan cheeses rely on mixtures of coconut oil and starches like tapioca, creating something good enough to put on a pizza, but not a cheese board. This story follows the rise and struggles of Climax Foods — rebranded as Bettani Farms in late 2025 — which tried to solve the problem using … AI. The company screened large numbers of plant compounds in search of combinations that could mimic the properties of dairy proteins, and in some respects it worked. The company developed a plant-derived ingredient called Caseed, which helped with the melt and stretch problem. But Bettani Farms was trying to expand as enthusiasm for other food startups, like Oatly and Beyond Meat, was waning, and Caseed proved expensive to manufacture. The company has restructured, cut products, ousted its CEO and rebranded to survive. Other researchers are now pursuing precision fermentation, using microbes to manufacture actual casein proteins without cows. In theory, that could produce vegan cheese that is chemically much closer to the real thing. But even if it’s technically possible, hitting the trifecta of making it affordable, scalable and genuinely tasty will be another challenge entirely.

Will We Ever Be Able To Forecast Volcanic Eruptions Like Weather? From Quanta
3,000 words, or about 12 minutes

Forecasting volcanic eruptions is hard. Scientists can tell when a volcano is becoming restless, but predicting exactly if, when and how it will erupt remains closer to educated guesswork than genuine prediction. A big issue is a lack of data about what are very complicated systems. Magma moves miles underground where it’s hard to study, through hidden networks that differ dramatically from one volcano to another. Each one has its own chemistry, pressure dynamics and geological quirks. That has made it difficult to create equations that describe volcanic behavior well enough to make reliable predictions. This is starting to change. At heavily monitored and regularly active volcanoes such as Mount Etna, Stromboli, Kīlauea, and those on Iceland’s Reykjanes Peninsula, scientists have become surprisingly good at forecasting eruptions hours before they happen. Scaling that capability across the globe will, according to this Quanta story, require more advanced monitoring across more volcanoes to collect the data needed for better predictive models. The hope is that those models would be able to ingest information about any volcano, then forecast eruption timing, intensity and duration. Such efforts are underway, including projects in the eastern Caribbean and across volcanic regions in Chile, Alaska and the US Cascades. The volcanologists working on these projects seem increasingly confident that they can discover the physical rules that could make eruption prediction a science that is more akin to weather prediction.