How the war in Iran threatens the global chip supply - Sync #564
Plus: Claude Mythos; OpenAI kills Sora; ARG-AGI-3; cryosleep gets closer to reality; Unitree IPO; TurboQuant; Amazon acquires humanoid robotics startup; Meta and YouTube lose landmark trial; and more!
Hello and welcome to Sync #564!
This week, we will take a look at how the war in Iran and the closure of the Strait of Hormuz is impacting global supply chains needed to produce the most advanced AI chips.
Over in AI, Anthropic confirmed it is working on new, powerful models and scored a victory over the Pentagon in court. Meanwhile, OpenAI shuts down the Sora app; Elon Musk launches his own chip factory; Arm unveils its own AI chip, Google releases TurboQuant and Lyria 3 Pro; and a Super Micro co-founder is arrested for reportedly smuggling advanced AI chips to China.
In robotics, Amazon buys a humanoid robotics startup, Wing expands to Bay Area, and Unitree prepares for an IPO.
Beyond that, this week’s issue of Sync also features Meta and YouTube being found negligent for designing addictive platforms, a conversation with Jensen Huang, scientists bringing cryosleep one step closer to reality, how computers are helping design new drugs, the OpenClaw boom in China, ARC-AGI-3 and the ARC Prize 2026, and more!
Enjoy!
How the war in Iran threatens the global chip supply
The Strait of Hormuz is one of the most important chokepoints in the global supply chain. Before the Iran war, roughly 20 million barrels of oil passed through it every day—about a quarter of the world’s seaborne crude. But oil is not the only product flowing through the strait. Approximately 20% of global LNG trade, a third of the world’s helium supply, and a steady flow of petrochemicals, fertilisers, and raw materials that feed the global economy also transit the waterway. When Iran closed the Strait on 2 March 2026, it severed one of the most critical supply arteries on earth.
The technology industry, which depends on that artery for everything from the energy powering chip factories to the helium cooling the wafers inside, is among the most exposed.
The Gulf as tech’s promised land
Less than a year ago, the Persian Gulf was the hottest destination in tech. During President Trump’s four-day tour of Saudi Arabia, Qatar, and the UAE in May 2025—flanked by executives from Nvidia, OpenAI, Oracle, Microsoft, Amazon, and Google—roughly $2.5 trillion in combined investment pledges were announced. AWS committed $5.3 billion to Saudi Arabia. Google Cloud entered a $10 billion joint venture with Saudi’s Public Investment Fund. The UAE unveiled Stargate UAE, a 5-gigawatt AI campus in Abu Dhabi backed by OpenAI, Oracle, and G42.
The logic behind those deals was straightforward. Gulf states had the capital and wanted cutting-edge AI built on their soil. In return, they got the technology and, implicitly, American protection.
Then the war put all of it in the line of fire.
The chip industry, powered by the Gulf
Oil prices have surged since the war began, with some analysts expecting Brent crude to reach $150 per barrel or even higher.
The countries hardest hit by the closure are those that rely on Gulf oil for energy production, which includes the places where most of the world’s semiconductors are made. Taiwan, home to TSMC, imports 97% of its energy and sourced roughly 70% of its crude oil and 38% of its natural gas from the Middle East before the crisis. South Korea, where Samsung and SK Hynix together produce 70% of global DRAM and 80% of high-bandwidth memory, depends on the Gulf for over 70% of crude imports. Japan, a major supplier of semiconductor equipment, sources 95% of its crude from the Middle East.
Then there is helium. Qatar produces roughly 35% of the world’s supply as a byproduct of LNG processing. When Iranian drones struck QatarEnergy’s Ras Laffan facility—the world’s largest LNG complex—and the company declared force majeure, helium production halted. Helium is an important resource in semiconductor manufacturing. It cools wafers during etching, shields extreme ultraviolet lithography systems, and enables plasma processes. No viable substitute exists. The prices surged 70-100%, and most fabs carry fewer than three months of inventory.
Markets responded accordingly. Samsung and SK Hynix—the two companies that make most of the world's memory chips—lost roughly $170 billion in combined market value over two trading sessions in early March, triggering emergency trading halts in South Korea. Memory chip prices, already elevated due to the high demand from the AI industry, are forecast to jump 55–60% this quarter alone.
The broken Gulf dream
The Gulf states were selling the dream of a luxurious, cosmopolitan, and safe oasis in the Arabian desert—global hubs for finance, tourism, tech, sports, and more. The billions they made along the way flowed back out as investments across the world, with AI infrastructure among the biggest bets. MGX, Abu Dhabi’s sovereign-backed AI investment fund, has become one of the most active technology investors in the world. It is a founding partner of OpenAI’s $500 billion Stargate project alongside SoftBank and Oracle, and has invested in Anthropic, xAI, and Databricks. Together with BlackRock and Microsoft, MGX launched a dedicated AI infrastructure fund. Saudi Arabia’s Public Investment Fund is backing a $10 billion data centre joint venture with Google Cloud.
When the missiles and drones hit, the Gulf dream shattered.
The $2.5 trillion in tech investment pledges from 2025 may not collapse outright, but renegotiations look very likely. The war puts at risk the AI buildout that the entire US economy is increasingly banking on—and Gulf money is central to that buildout. It is costing Gulf states billions of dollars per day in lost oil revenue, grounded flights, and disrupted trade. It will take years to rebuild the image they held before the war (if that is even possible) which will add further billions to the total cost. Those are resources that would otherwise have gone into AI infrastructure and data centres. If Gulf sovereign wealth funds decide the geopolitical bet was miscalculated and redirect capital, the downstream effects would hit AI funding rounds, data centre construction timelines, and hyperscaler expansion plans worldwide. If that were to happen, AI sceptics may have finally found the pin that pops the AI bubble.
The ripple effects are coming
The Iran war has already sent shockwaves across the global economy. But like a tsunami after an earthquake, the bigger waves are on the way.
The semiconductor production price index historically correlates with oil prices after a six-month lag—meaning March’s surge will not fully manifest in chip costs until September. Helium reserves at most fabs will not last beyond May if the Strait remains closed. If those stockpiles run dry, production at some of the world’s most advanced chip factories could halt entirely. Even a swift resolution would not undo the damage— it may take a minimum 4–6 month for helium supply chain to return to normal after the conflict ends.
The industry that promised to make the world frictionless, borderless, and virtual remains tightly coupled with tanker routes, gas pipelines, and a narrow strait between Iran and Oman. The consequences of the war in Iran will run well into 2027, even if the bombs and missiles stop falling tomorrow.
If you enjoy this post, please click the ❤️ button and share it.
🦾 More than a human
Reviving Brain Activity After ‘Cryosleep’ Inches Closer in Pioneering Study
Researchers froze mouse brain tissue into a glass-like state to avoid the ice crystal damage caused by conventional freezing, then thawed it after up to a week in cryogenic storage. The neurons resumed electrical activity and showed signs of the processes involved in learning and memory. The technique worked on both brain slices and whole brains, though results varied across different neuron types. The team is now extending the work to human brain tissue. While true memory preservation remains unproven, the findings suggest brains are far more resilient to deep freezing than previously believed.
🔮 Future visions
▶️ Biohacks & Brain Mods - The Coming Age of Implant Culture (28:36)
Isaac Arthur looks at how brain implants might develop, from today's cochlear implants and brain-computer interfaces to future memory aids, mood regulators, and eventually full neural meshes. He considers the complications, from subscription pricing, black markets, widening inequality, pressure to upgrade just to stay competitive, to the need to protect not just data but thoughts from hacking. Despite these concerns, he remains optimistic and believes humanity will learn to handle neurotech much as it has with every major advance before it.
🧠 Artificial Intelligence
Judge blocks Pentagon’s effort to ‘punish’ Anthropic by labeling it a supply chain risk
A federal judge has blocked the Pentagon from designating Anthropic a supply chain risk, ruling it was unconstitutional retaliation for the company’s refusal to allow its AI to be used in autonomous weapons or mass surveillance. The judge called it “classic illegal First Amendment retaliation,” noting the label had previously only been applied to foreign adversaries. The Pentagon plans to appeal.
Anthropic Denies It Could Sabotage AI Tools During War
The US Department of Defense claims Anthropic could disable or tamper with Claude during active military operations, but the company flatly denies this, stating in court filings that it has no back door, kill switch, or remote access to deployed models. The dispute escalated when defence secretary Pete Hegseth designated Anthropic a supply-chain risk, barring the Pentagon and its contractors from using Claude. Anthropic has filed two lawsuits challenging the ban and offered contractual guarantees relinquishing any veto over military decisions, though negotiations broke down.
Anthropic acknowledges testing new AI model representing ‘step change’ in capabilities, after accidental data leak reveals its existence
Anthropic has confirmed it is building its most powerful AI model yet, called Claude Mythos, after a data leak exposed its existence. The leak left thousands of unpublished files publicly accessible until Fortune flagged the issue. Anthropic says the model far outperforms its current best in coding, reasoning, and cybersecurity, but warns it also poses serious risks, as hackers could use it to find and exploit software weaknesses faster than defenders can fix them. The company is rolling the model out cautiously, starting with a small group of early testers.
Musk Says Tesla, SpaceX, xAI Chip Project to Kick Off in Texas
Elon Musk launches Terafab, a chip manufacturing project in Austin, Texas, jointly operated by Tesla and SpaceX. The facility will produce two types of chips—one optimised for edge and inference tasks in vehicles and robots, and another high-power chip for space computing by SpaceX and xAI. Musk argues the semiconductor industry cannot scale fast enough to meet his companies’ AI and robotics demands, though he offered no timelines and has a history of overpromising. The project represents an enormously costly bet on vertical integration in an industry where new fabs typically take years and tens of billions of dollars to bring online.
OpenAI shuts down Sora
OpenAI is shutting down Sora, its AI video generator app. This move is part of OpenAI’s strategy to scale back side projects and focus on coding tools and enterprise customers—more profitable markets currently dominated by Anthropic. Sora’s shutdown has also made Disney exit its $1 billion deal with OpenAI.
OpenAI Plans Launch of Desktop ‘Superapp’ to Refocus, Simplify User Experience
The Wall Street Journal reports that OpenAI plans to merge ChatGPT, its Codex coding platform, and the Atlas browser into a single desktop “superapp”, reversing last year’s strategy of launching fragmented standalone products. The superapp will prioritise agentic AI capabilities that can autonomously perform tasks such as writing code and analysing data, which make it sound similar to Claude Cowork.
The Silicon Valley Salesman Accused of Helping China Get Nvidia’s Top Chips
Super Micro co-founder Wally Liaw was arrested for allegedly helping smuggle $2.5 billion worth of Nvidia AI chip servers to China, using shell companies and staged warehouse inspections to evade US export controls. The charges cast doubt on assurances from Nvidia and the White House that chip smuggling is not a serious problem, and sent Super Micro’s share price down by a third, wiping out over $6 billion in value.
Announcing Arm AGI CPU: The silicon foundation for the agentic AI cloud era
Arm is joining the AI chip game with AGI CPU, its first in-house silicon product and a significant shift from its traditional IP licensing model. The chip targets agentic AI workloads that demand sustained, massively parallel CPU performance. Arm claims over 2x rack-level performance compared to x86 competitors. Meta has been announced as the lead customer, with OpenAI, Cerebras, Cloudflare, and others signed on as launch partners. Production is expected later in 2026.
Put Claude to work on your computer
Claude Cowork and Claude Code can now point, click, and navigate user’s computer to complete tasks. The new feature is available in research preview for Claude Pro and Max subscribers. Additionally, Anthropic introduced auto mode, a new permissions mode in Claude Code where Claude makes permission decisions on user’s behalf, with safeguards monitoring actions before they run.
TurboQuant: Redefining AI efficiency with extreme compression
Google has announced TurboQuant, a new compression algorithm for LLMs and vector search. According to Google, it compresses key-value caches to just 3 bits with no accuracy loss and up to 8x speedup on H100 GPUs. TurboQuant combines two novel techniques: QJL, a 1-bit method eliminating memory overhead, and PolarQuant, which uses polar coordinates to avoid costly normalisation. Google says the methods are theoretically grounded and plans to apply them to Gemini and large-scale semantic search. Internet, meanwhile, noticed some similarities between TurboQuant and Pied Piper.
ARC-AGI-3 and ARC Prize 2026
ARC Prize Foundation launched ARC-AGI-3, an interactive benchmark that tests whether AI agents can explore novel environments and learn from experience as efficiently as humans. Rather than measuring static puzzle-solving, it evaluates skill acquisition, long-horizon planning, and real-time adaptation over multiple steps. Currently, the best models in the world score 0%, while humans can easily solve them. You can try the puzzles yourself here. The foundation also launched ARC Prize 2026, a $2 million competition across three tracks rewarding top open-source solutions for ARC-AGI-3 and ARC-AGI-2 as well as research papers advancing AGI understanding of how to achieve strong performance on ARC-AGI.
Bernie Sanders and AOC propose a ban on data center construction
Bernie Sanders and Alexandra Ocasio-Cortez have proposed banning new data centres over 20 megawatts until Congress passes comprehensive AI regulation, citing widespread public concern and warnings from leading technologists. The broader bill also seeks government certification of AI models before release, worker protections, environmental limits, and chip export controls. Passage is unlikely given industry lobbying and competition with China, but the bill signals an opening move in the AI governance debate.
MiniMax M2.7: Early Echoes of Self-Evolution
MiniMax has released M2.7, a large language model that can help build and improve itself—creating its own tools, refining its setup, and running reinforcement learning experiments with minimal human input. Independent benchmarks from Artificial Analysis confirm it is one of the best open models, behind only top Western models like Gemini 3.1 Pro, GPT-5.4, Claude Opus 4.6, and Claude Sonnet 4.6. However, the same evaluation found M2.7 to be slow, overly wordy, and somewhat expensive compared to similarly priced rivals. If you want to learn more about M2.7, and specifically about the claims about self-evolution, here is a good video about MiniMax’s latest model.
Google has released Lyria 3 Pro, an improved AI tool that can generate music tracks up to three minutes long, with users able to specify song structure like verses and choruses. It's being made available across several Google products for businesses, developers, and everyday creators. Google says the tool was built with input from music professionals and includes safeguards such as watermarking to identify AI-generated content.
China’s OpenClaw Boom Is a Gold Rush for AI Companies
OpenClaw is going big in China, and the real winners aren't the users—they're the tech companies. The open-source AI tool went viral after influencers showed it off automating tasks like stock trading, but ordinary people quickly found it too technical to set up or use without programming knowledge. Meanwhile, companies like Tencent, ByteDance, and Moonshot have cashed in on the hype, selling cloud services and API access while releasing their own versions to keep users on their platforms.
Mark Zuckerberg Is Building an AI Agent to Help Him Be CEO
Mark Zuckerberg is reportedly building a personal AI assistant to help him do his job faster, and he wants the rest of Meta to follow suit. I wonder if Zuckerberg’s AI agent will be good enough to push back on some ideas or if it will make him delusional.
Palantir Demos Show How the Military Could Use AI Chatbots to Generate War Plans
A WIRED investigation reveals how Anthropic's Claude AI is being used in US military operations through Palantir's software, helping analysts review intelligence, suggest attack plans, and support decisions about strikes. This comes as Anthropic and the Pentagon are locked in a legal fight over the company's refusal to give the military unrestricted access to Claude. Neither company has said which military systems use Claude or how it is being used in current operations.
Leaders of AI Firm Bought by Meta Are Restricted From Leaving China
China has barred two co-founders of AI startup Manus from leaving the country while regulators review its $2.5 billion sale to Meta. The Chinese government is concerned that the company relocated operations from Beijing to Singapore to sidestep oversight of cross-border technology transfers. Authorities are investigating whether ownership changes were properly reported and are considering penalties, while Meta insists the deal was lawful and that Manus no longer has any Chinese ties.
Cursor admits its new coding model was built on top of Moonshot AI’s Kimi
Cursor faced backlash after an X user revealed its new Composer 2 model was built on Kimi 2.5, an open-source model from Chinese company Moonshot AI. Cursor confirmed the open-source base but said three-quarters of the final model's compute came from its own training. The incident underscores the sensitivity around US companies quietly building on top of Chinese AI models, and Cursor's co-founder acknowledged it was "a miss" not to credit Kimi upfront.
In this wide-ranging conversation, Jensen Huang discusses how he runs the company and explains Nvidia’s engineering approach of designing chips, networking, memory, cooling, and software as one tightly integrated system. On the future of AI, Huang outlines multiple ways intelligence will keep scaling. The conversation also covers Nvidia’s relationship with TSMC, China's innovation ecosystem, Nvidia’s competitive moats, the potential for AI agents to create billion-dollar businesses, supply chain challenges, and more.
What 81,000 people want from AI
Anthropic interviewed over 80,000 Claude users across 159 countries and asked them what they want and worry about with AI. Most people wanted AI to give them a better life rather than just help them work faster, and 81% said it was already making a difference. The main finding is that the same people who benefit most from a particular use of AI are often the ones most worried about its downsides. People in lower-income countries tended to be more positive, seeing AI as a way to access new opportunities.
🤖 Robotics
Amazon acquires humanoid developer Fauna Robotics
Amazon has acquired Fauna Robotics, a humanoid robot startup that recently came out of stealth, which will join its Personal Robotics Group. Fauna's product, Sprout, is a small, safety-focused humanoid currently aimed at researchers, but Amazon has hinted at broader personal robotics ambitions.
Wing’s homecoming: Bringing drone delivery to the Bay Area
Wing, Google’ drone delivery company, annouced that it is expanding its service to the San Francisco Bay Area in the coming months. The company has completed over 750,000 deliveries across US cities like Houston, Atlanta, and Dallas, partnering with brands such as Walmart and DoorDash, and sees the Bay Area launch as part of its push to build a nationwide drone logistics network.
Unitree IPO shows a real hardware business, but the humanoid case is still early
Unitree Robotics has filed for a STAR Market IPO in Shanghai, aiming to raise roughly $610 million on reported 2025 revenue of about $248 million. The company has grown quickly, with humanoid robots now making up over half its sales—up from almost nothing two years ago. Interestingly, it has slashed prices from around $85,000 to $25,000 per robot while still making healthy profits, suggesting it has found ways to build them much more cheaply. The catch is that most buyers are still universities and researchers, not factories, so while Unitree looks like a solid hardware business, the bigger promise of robots working widely in industry hasn't been proven yet.
Zipline snaps up another $200M to fuel its drone delivery expansion
Zipline has raised an additional $200 million, bringing its Series H round to $800 million at a $7.6 billion valuation. The drone delivery company, which started delivering blood in Rwanda in 2014, now operates across five African countries, the US, and Japan. The funding will accelerate expansion into at least four new US states, where its home delivery service is exceeding growth forecasts, while a new national-scale contract in Rwanda will extend its reach across the country's cities and healthcare facilities.
▶️ Meet “Roadrunner”: a bipedal, wheeled robot for multi-modal locomotion (1:33)
Researchers at RAI Institute, a robotics research lab spun off from Boston Dynamics, present Roadrunner, a new bipedal wheeled robot prototype designed to seamlessly switch between its side-by-side and in-line wheel modes, as well as stepping configurations, depending on the requirements of its environment. The robot’s legs are entirely symmetrical, allowing it to point its knees forwards or backwards, which can be used to avoid obstacles or manage specific movements.
A former Thiel fellow’s startup just launched a drone it says can replace police helicopters
Brinc, a US drone startup valued at nearly half a billion dollars, has launched Guardian—a public safety drone equipped with embedded Starlink connectivity, thermal and 4K cameras, and an automated charging station. The drone can also deploy emergency supplies, such as defibrillators. The company estimates a $6-8 billion market for equipping police and fire stations with rooftop 911 response drones. This ambition has been boosted by the Trump administration’s recent ban on foreign-made drones, which has weakened the dominance of China’s DJI.
▶️ Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data (1:31)
Researchers from China have taught a humanoid robot to play tennis by training it on rough snippets of human movement rather than needing perfect recordings of real matches. The robot, a Unitree G1, can move around the court, hit the ball back, and even keep a rally going with a human player. It's far from competing with any real tennis player, but it's an impressive step forward for getting robots to handle fast, physical tasks in the real world.
🧬 Biotechnology
▶️ Here’s how computers are helping build antibiotics (20:50)
In this video, Clockwork (an excellent YouTube channel on biochemistry) explains how machine learning is transforming antibiotic discovery, using Zosurabalpin a new antibiotic from Roche now entering phase 3 trials—as its central example. The compound was identified by feeding experimental results into a random forest algorithm (no generative AI involved) to rapidly narrow the field to a handful of viable candidates. The video also highlights newer work from MIT using generative AI to screen millions of chemical fragments into promising compounds. Despite cautioning that real-world results remain years away, the video argues we are witnessing a genuinely transformative moment in drug discovery.
Trillion Gene Atlas Expands Evolutionary Datasets for Next-Generation AI Therapeutics
Basecamp Research has teamed up with Anthropic, Nvidia, and others to launch the Trillion Gene Atlas, a project aiming to sequence DNA from over 100 million species to build a massive training dataset for AI drug discovery. Current models are limited by narrow public datasets, but Basecamp's proprietary data—already ten times larger—has shown promising results in designing therapies. The goal is to let AI design medicines directly from a disease description, with Nvidia’s computing power compressing decades of processing into under two years.
Scaling frontier models in vitro
JURA has built an AI platform that designs its own biological experiments and learns from the results in a continuous loop. In a three-day demonstration, it tested over 209 million antibody candidates against 100 cancer-related targets and trained a model capable of predicting successful matches even for entirely new targets. The company argues this tight integration of AI and lab work opens up a fundamentally new approach to drug design.
▶️ Why Building New Proteins from Scratch Is Our New Superpower | David Baker (30:40)
In this TED interview, Nobel laureate David Baker discusses his pioneering work in computational protein design. Since his 2019 TED talk, applications have expanded dramatically—from an approved COVID vaccine to enzymes that degrade plastics and forever chemicals, heat-tolerant crop proteins, and biosensors that interface with electronics. Baker advocates strongly for open-source science and argues that biosecurity risks remain low compared to the benefits, since engineering dangerous pathogens is far more complex than designing single-function proteins.
💡Tangents
Jury finds Elon Musk misled investors during Twitter purchase, absolves him of some fraud claims
A jury found Elon Musk liable for misleading Twitter investors with two tweets that drove down the company's share price during his 2022 takeover bid, awarding shareholders roughly $2.6 billion in damages. However, jurors stopped short of finding he deliberately schemed to defraud investors. Musk's legal team plans to appeal, while plaintiffs' lawyers hailed the verdict as a signal that market-moving statements carry legal consequences regardless of the speaker's wealth.
Meta and YouTube Lose Landmark Social-Media Addiction Trial
A Los Angeles jury found Meta and YouTube negligent for designing addictive platforms, awarding $6 million to a woman who said social media use from childhood led to anxiety, depression, and poor body image. The case got around existing legal shields by focusing on how the apps were designed rather than what users posted on them. With over 3,000 similar cases waiting to be heard and a second ruling against Meta in the same week, the verdict marks a turning point in holding social media companies responsible for their effects on children.
▶️ How to build your dragon (46:42)
If you’ve ever wondered how to create a real-life dragon, this video is for you. The Thought Emporium, YouTube’s chief mad scientist, explains how a real dragon might work and how one could be created—if genetic engineering were advanced enough.
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!"