OpenAI crashed global memory supply - Sync #546

Plus: Gemini 3 Pro; Anthropic reaches $350B valuation; bad week for tech stocks; Grok 4.1; GPT-5.1-Codex-Max; Bezos is back with AI startup; Waymo and Zoox expand services; the future of humanity

Hello and welcome to Sync #546!

This week’s main story is about computer memory and how OpenAI effectively crashed that industry. We’ll explore how one deal sent memory chip prices soaring, and how it may impact the wider tech sector and consumers.

Elsewhere in AI, Google has released Gemini 3 Pro, OpenAI has launched GPT-5.1-Codex-Max, and xAI has introduced Grok 4.1. Meanwhile, Anthropic has reached a $350 billion valuation, Jeff Bezos is back with a new AI start-up, and the entire tech sector has had another bad week on Wall Street.

Over in robotics, Waymo and Zoox are expanding their services, Figure has shared what it learned from its partnership with BMW, and Disney is teaching small robots to fall with style.

In other news, CRISPR has supercharged a meat-like fungus into a sustainable source of protein; Meta has won its antitrust trial; the first nuclear start-up has achieved criticality; what humanity could look like in the near and far future; and more!

Enjoy!

OpenAI crashed global memory supply

Photo by Liam Briese on Unsplash

The AI boom and the rush to build AI infrastructure have pushed the prices of computer hardware sharply upward. GPUs were already expensive, and now, thanks in large part to OpenAI, memory is becoming expensive too. OpenAI’s massive wafer deal with Samsung and SK hynix has effectively crashed the global DRAM market, triggering shortages that are rippling through the tech industry and pushing prices higher for everyone.

Why memory prices are exploding

At the beginning of October 2025, OpenAI announced a new partnership with Samsung and SK hynix, the two largest memory suppliers. The partnership, part of OpenAI’s Stargate initiative, will see OpenAI securing 900,000 DRAM wafers per month, or 40% of global DRAM output.

What made the situation so destabilising for the industry wasn’t just the scale of the deal, but how suddenly it landed. According to industry reporting, including insider accounts shared by Moore’s Law Is Dead, Samsung and SK hynix themselves allegedly did not know the scale or timing of each other’s agreements until the day they were announced. The industry effectively woke up to find that almost half of global DRAM output had been suddenly locked down. The secrecy triggered an immediate wave of panic buying, as every company reliant on DRAM—from cloud providers to laptop manufacturers—rushed to secure whatever inventory remained.

This scramble collided with an unusually fragile supply chain. DRAM prices had been trending downward earlier in the year, discouraging stockpiling. Unpredictable tariffs made companies hesitant to build large inventories. Manufacturers had also spent two years cutting output to recover from the COVID-era overstock collapse, leaving safety stock thinner than usual. When OpenAI’s announcement landed, it hit a supply system with almost no buffer left. Warehouses emptied, wholesalers burned through allocations, and retailers began reporting fulfilment timelines stretching well into 2026.

In response, DRAM prices rose by 40–50% within weeks, in some cases even faster. Some retailers were told that orders placed now might not be fulfilled until late 2025 or even 2026. Industry insiders reported extraordinary behaviour, such as memory brands attempting to buy back their own modules from retailers to satisfy other contracts.

How much has the price increased can be seen on PCPartPicker. It tracks the prices of various computer components, including memory. The prices for memory has been raising for some time now, but in the last two months it has skyrocketed in some cases. GamersNexus found DRAM kits climbing by 100–180% in one to two months, including examples where prices rose nearly 1% per day for almost two weeks.

The memory crunch is here—what you should know

At the moment, the memory is the most impacted, but soon the fallout will spread to other components and devices.

Desktop DRAM are the first to spike in price, hitting PC builders and DIY users hardest. SSDs, which use DRAM chips to improve performance, will follow suit. Graphics cards, which are already expensive, will come under increasing pressure as manufacturers struggle to secure sufficient memory. AMD is likely to be affected more severely than Nvidia due to smaller stockpiles. Some new GPU models, particularly those planned with higher memory capacities, may be pushed back or cancelled altogether.

In the gaming market, Xbox consoles could face supply constraints owing to limited memory buffers, while PlayStation systems should remain better insulated after making earlier bulk purchases. Laptops and smartphones from major brands should continue shipping without immediate disruption, but analysts expect rising costs to reach them by mid-2026. CPUs, meanwhile, are expected to stay largely untouched, as they are one of the few classes of hardware not directly constrained by DRAM availability.

Why OpenAI bought raw wafers, not chips

Moore’s Law is Dead also highlights an intriguing thing about OpenAI’s deal with memory manufactures. Apparently, OpenAI did not secure completed DRAM chips or packaged HBM stacks ready for use in servers. Instead, Samsung and SK hynix will supply undiced wafers—unfinished memory that still needs to be cut, tested, packaged, and assembled. This is highly atypical for a customer, and suggests that OpenAI’s priority is not immediate deployment but long-term control.

By locking in raw wafer capacity rather than finished products, OpenAI ensures that competitors cannot access the same supply, even if it takes years to convert those wafers into usable memory. In a market where HBM and server DDR5 already consume far more wafer capacity than consumer RAM, this strategy effectively starves rivals of a resource as fundamental to AI scaling as electricity itself.

A bleak outlook for 2026

Looking ahead, the outlook for consumers is bleak. Analysts expect memory prices to remain elevated throughout 2026, with some forecasting a multi-year shortage as AI servers consume an ever-greater share of global wafer capacity. There is also a possibility of a “structural memory price surge”, driven by cloud providers outbidding every other sector for HBM and server-grade DDR5, while manufacturers prioritise these high-margin products over desktop and mobile components. NAND flash is expected to follow the same trajectory: SSDs could see further 50–100% price rises as data-centre storage demand intensifies. The traditional three-to-four-year DRAM pricing cycle may be breaking entirely, replaced by a prolonged period of heightened prices, wafer scarcity, and supply shifted towards AI infrastructure at the expense of the consumer market. Unless the AI investment boom cools suddenly, the memory market is likely to remain in shortage throughout 2026 and maybe even beyond.

Unless there is a sharp reversal in AI spending, 2026 may mark the beginning of a new normal: memory as a permanently scarce commodity, just like GPUs have been for the past few years. With cloud giants now competing directly with consumers for every die, and with wafer production stretched to its limits, the era of cheap DRAM and SSDs could be over. What happens next will depend less on technology and more on whether the AI bubble deflates. For now, AI data centres are swallowing the world’s supply, and consumers are left fighting for scraps.

---

If you want to learn more about this topic, I highly recommend checking out the recent videos from Moore’s Law Is Dead and GamersNexus. Both channels offer clear, detailed explanations of what’s happening inside the memory industry, and are well worth your time and attention.

If you enjoy this post, please click the ❤️ button and share it.

---

🦾 More than a human

Paradromics Gets FDA Approval to Trial Its Brain Implant in People
Paradromics has received FDA approval to begin a small human trial of its brain implant, which is designed to help people who can no longer speak due to severe motor impairment. The device reads brain signals linked to attempts to speak and turns them into text or synthesised speech, possibly using a cloned version of the user’s own voice. The trial will start with two people next year and may expand later.

This Wireless Brain Implant Is Smaller Than a Grain of Salt
MOTE is a tiny, wireless brain implant powered and controlled by light that can record neural activity for over a year with minimal tissue damage. Smaller than a grain of salt, it converts red and infrared light into energy and transmits data through light pulses, allowing it to move with the brain and avoid scarring common with traditional wired implants. MOTE was successfully tested in mice, where it captured both single-neuron and network activity. This device could one day be used in organoids or other tissues, offering a durable and less invasive way to study the brain, and maybe even find use in brain-computer interfaces.

🔮 Future visions

▶️ The Future of Humanity - What Will We Become? (29:03)

In this video, Isaac Arthur asks what humanity might become across short, long, and cosmic timescales. He explores how life extension, AI integration, and physical and mental augmentation may transform us from baseline humans into diverse transhuman and posthuman forms. He considers how these changes could reshape societies, drive human expansion into the Solar System and beyond, and ultimately branch into many descendant civilisations that endure across billions or even trillions of years.

🧠 Artificial Intelligence

A new era of intelligence with Gemini 3
Google has released its long-awaited new flagship model, Gemini 3 Pro, its smartest yet. It shows strong results across reasoning, image/video understanding and coding benchmarks. The model is already being rolled out in Google products, with the Deep Think mode planned for later. You can read more about Gemini 3 Pro in my post here.

Gemini 3 Pro puts Google back on top

Tech stocks wrap big losing week as AI names get rocked after Nvidia earnings
Even Nvidia’s great results and CEO Jensen Huang’s rejection of an “AI bubble” couldn’t prevent a broad sell-off. Nvidia reported revenue of $57 billion in the third quarter, 62% higher compared to the same quarter last year. Its shares initially rose more than 5% before reversing to a 3.2% decline. Most of the Magnificent Seven also fell, with Amazon and Microsoft hit the hardest, while Alphabet was the only major gainer thanks to the Gemini 3 launch. Nvidia-linked firms such as Oracle, CoreWeave and several chipmakers also dropped sharply.

Anthropic valued in range of $350 billion following investment deal with Microsoft, Nvidia
Anthropic has secured major new partnerships with Microsoft and Nvidia, with Microsoft set to invest up to $5 billion and Nvidia up to $10 billion, bringing the AI startup’s valuation to around $350 billion. These agreements also reflect Microsoft’s efforts to broaden its AI partnerships beyond OpenAI, as Anthropic commits to significant Azure and Nvidia compute purchases. Additionally, Anthropic will collaborate with Nvidia on optimising future AI models.

Building more with GPT-5.1-Codex-Max
OpenAI has launched GPT-5.1-Codex-Max, a new coding model built to handle long, complex software tasks more efficiently. OpenAI says it can work across very large projects, run for hours or days if needed, and performs better at tasks like code review and debugging. GPT‑5.1-Codex-Max is available in Codex with ChatGPT Plus, Pro, Business, Edu, and Enterprise plans, with API access on the way.

Elon Musk’s xAI Is in Advanced Talks to Raise $15 Billion, Lifting Valuation
xAI is reportedly close to raising $15 billion at a valuation of around $230 billion. The company is rapidly spending money to build large AI infrastructure and recently raised $10 billion for its Colossus data centre. Musk publicly denied an earlier report about the new fundraising, and the company has seen several senior executives leave.

Grok 4.1
xAI has released Grok 4.1, an updated version of its AI model, which comes in two variants: thinking and non-thinking. According to the company, the new models topped the LMArena Text Leaderboard and EQ-Bench, and were only second to GPT-5.1 on the Creative Writing v3 benchmark at the time of their release. In an interview, Elon Musk said xAI is planning to release Grok 5 in Q1 2026.

Jeff Bezos reportedly returns to the trenches as co-CEO of new AI startup, Project Prometheus
Jeff Bezos is back. Amazon’s founder returns as co-CEO of a new startup named Project Prometheus. The company, which already employs nearly 100 staff, including researchers from major AI labs, has raised $6.2 billion in funding and aims to build technology that simulates the physical world to accelerate scientific and industrial innovation. Bezos will share the leadership role with Vik Bajaj, a veteran of Google’s life sciences efforts and the co-founder of several biotech and AI ventures.

Foxconn, OpenAI partner on AI hardware manufacturing
Foxconn and OpenAI are teaming up to design data centre hardware for artificial intelligence, giving OpenAI early access to test the new systems while helping Foxconn better understand the needs of major AI companies. Foxconn will build the components in the United States to shore up supply chains and avoid possible tariffs.

Larry Summers resigns from OpenAI board after release of emails with Epstein
Former Treasury Secretary Larry Summers has stepped back from teaching at Harvard and resigned from the board of OpenAI after emails between him and sex offender Jeffrey Epstein were released. Summers has apologised, saying he is ashamed of his past communications with Epstein.

a16z-backed super PAC is targeting Alex Bores, sponsor of New York’s AI safety bill — he says bring it on
A new pro-AI super PAC, Leading the Future—backed by major tech figures including Andreessen Horowitz and OpenAI’s Greg Brockman—has chosen New York Assembly member Alex Bores as its first target because of his support for stronger state-level AI regulation. Bores is promoting the RAISE Act, which would require large AI firms to follow safety plans and avoid releasing high-risk models. Tech backers of the PAC argue the bill would harm US innovation and want a single national framework that limits state action, while Bores says states must step in because Congress has not passed any meaningful AI laws.

Sakana AI raises $135M Series B at a $2.65B valuation to continue building AI models for Japan
Tokyo-based Sakana AI has raised ¥20 billion (about $135 million) in a Series B round, lifting its valuation to $2.65 billion. Founded in 2023 by former Google researchers, the company plans to expand its R&D and workforce, and says it will focus on building efficient, Japan-optimised models through sustainable post-training rather than large-scale model building.

WeatherNext 2
Google DeepMind and Google Research have launched WeatherNext 2, their most advanced weather-forecasting model yet, capable of generating hundreds of possible scenarios in under a minute with high resolution and improved accuracy across temperature, wind, humidity and pressure. According to the press release, the new model surpasses the previous state-of-the-art WeatherNext model on 99.9% of variables and lead times (0-15 days), enabling more useful and accurate forecasts. WeatherNext 2 is now available via Earth Engine, BigQuery and in early access on Vertex AI, and will power upgraded weather insights in Search, Maps and other Google services.

Microsoft to Use OpenAI’s Custom Chip Work to Help In-House Effort
Microsoft will use OpenAI’s custom AI chip and hardware work to strengthen its own chip efforts, Satya Nadella said. Their updated agreement gives Microsoft access to OpenAI’s models through 2032 and its research until 2030, excluding consumer hardware. OpenAI is designing chips with Broadcom, and Microsoft—behind rivals like Google in chipmaking—plans to build on both OpenAI’s designs and its own using the IP rights it now holds.

🤖 Robotics

Waymo is launching driverless robotaxis in 5 more cities
Waymo has begun fully autonomous driving in Miami and will roll out the service in Dallas, Houston, San Antonio and Orlando in the coming weeks, ahead of opening to riders next year. The company says its self-driving system is safer than human drivers and is improving as it expands to new cities. Waymo adds that it works closely with local officials and communities to build trust and ensure its autonomous ride-hailing service is safe and reliable.

Zoox offers its first public autonomous rides in San Francisco
Amazon-owned Zoox has started its Explorers programme in San Francisco, offering free rides in its fully driverless, purpose-built robotaxis to people on its waitlist. Riders can travel around key neighbourhoods and share feedback to help improve the service. Zoox is expanding after launching in Las Vegas and testing in several other cities.

F.02 Contributed to the Production of 30,000 Cars at BMW
Figure shares in this post the lessons it learned from deploying its 02 humanoid robots at the BMW factory in Spartanburg, South Carolina. The company reports that over 11 months, the robots ran 10-hour shifts from Monday to Friday, logging more than 1,250 hours of runtime and helping to assemble over 30,000 BMW X3 cars. Figure says that the data gathered on speed, precision and reliability during this pilot highlighted issues such as forearm hardware failures and informed major design improvements in the next-generation Figure 03 robot. Notably, Figure does not mention whether its partnership with BMW has been extended.

A $2.6 billion humanoid robotics startup under the spotlight - Sync #523

▶️ A Dynamic Robot That Can Throw, Catch, and Hit a Baseball (0:54)

Researchers from the RAI Institute, a robotics laboratory spun off from Boston Dynamics, present in this video a robot that can throw, catch, and hit a baseball. The robot is capable of throwing a ball at 70 mph (112 kph) and can also catch and bat at short distances. The researchers highlight its quick reaction times, which allow it to catch and strike fast-travelling balls. Although the robot performs well at baseball, its primary purpose is to serve as a low-impedance platform for studying dynamic robotic manipulation.

Why Is Everyone’s Robot Folding Clothes?
If you’ve seen recent videos of robots in action, you may have noticed that many of them are folding clothes to show what they can do. If you’ve ever wondered why, this article explains why folding laundry is a popular choice for robotics demos.

▶️ Robot Crash Course: Learning Soft and Stylized Fallin (3:00)

Every humanoid robot will sooner or later fall, which can risk damaging the robot. In this video, researchers from Disney present the results of their work on training a small humanoid robot to fall softly and with style. Compared with traditional strategies, this new approach results in a gentler landing, thus reducing the likelihood of damage. Additionally, with this method, researchers can train the robot to protect specific components, such as batteries, during falls.

Magnetically Guided Microrobots Deliver Drugs with Pinpoint Accuracy
Researchers at ETH Zurich have created a magnetic microrobot system that can travel through the body’s vessels to deliver drugs directly to where they’re needed, reducing side effects, improving precision and offering promising potential for treating strokes, infections, and tumours. The tiny capsule—made from dissolvable, FDA-approved materials—can be steered using magnetic fields and releases its payload when heated. Tests in realistic vessel models and in animals showed that the system can navigate even very small or complex pathways and correctly deliver drugs in over 95% of cases.

▶️ Walker S2 - World’s First Mass Delivery: Behind the Scenes (1:17)

Almost two weeks ago, UBTECH, a Chinese robotics company, released a video showing a small army of its Walker S2 preparing to be shipped. The footage looked as though it was CGI or taken straight from a sci-fi movie. In response to these claims, UBTECH released a behind-the-scenes video to prove that the original footage was real. I think they could do a better job with that.

🧬 Biotechnology

Scientists Race to Deliver Custom Gene Therapies for Incurable Diseases in Weeks—Not Years
Scientists are accelerating the development of personalised gene editing therapies using base editing, a CRISPR-related tool that can fix single-letter DNA mutations behind rare metabolic diseases. Building on a recent successful case, doctors plan a trial in which several children will receive customised treatments that differ only in their guide RNA, and regulators have agreed to assess them under one streamlined safety process. This new approach could greatly cut the time and cost of creating tailor-made therapies and pave the way for personalised genetic treatments to become routine for rare inherited disorders.

CRISPR Supercharges a Meatlike Fungus Into a Sustainable Protein Powerhouse
Scientists used CRISPR to improve a fungus that naturally tastes and feels like meat, helping it grow faster, use fewer resources, and become easier to digest. By removing two genes, they made the fungus produce the same amount of protein with far less sugar and in much less time. The improved strain can cut greenhouse gas emissions by up to 60% and uses far less land and water than chicken farming, making it a promising, environmentally friendly protein alternative.

💡Tangents

Meta wins antitrust trial as judge denies that it’s a monopoly
A US judge has ruled in Meta’s favour in the Federal Trade Commission’s antitrust case, finding that regulators failed to prove Meta currently operates as a monopoly despite its past concerns about Instagram and WhatsApp as rising competitors. Judge James Boasberg noted that the social media landscape has shifted significantly since the lawsuit was filed, with strong rivals like TikTok showing that Meta still faces substantial competition.

Valar Atomics Says It’s the First Nuclear Startup to Achieve Criticality
Valar Atomics has become the first startup in a new US Department of Energy programme to achieve cold criticality, meaning its experimental reactor can sustain a controlled chain reaction. The test, done with Los Alamos National Laboratory, confirms the reactor’s design but does not generate power. The programme lets startups move faster than usual by skipping some early regulatory steps, though Valar will still need full approval before any commercial reactor can be built.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it.

Humanity Redefined sheds light on the bleeding edge of technology and how advancements in AI, robotics, and biotech can usher in abundance, expand humanity's horizons, and redefine what it means to be human.

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

My DMs are open to all subscribers. Feel free to drop me a message, share feedback, or just say "hi!"