Headlines this week - May 4, 2025
A look at how capital is being deployed across future opportunities
1 - Could muons make “cold” nuclear fusion possible? Substituting electrons with muons to orbit hydrogen or helium atoms’ nuclei makes it easier to force these nuclei to “merge” and trigger a nuclear fusion reaction, even at room temperatures (with no need to create and confine ultra-high temperature plasma).
This has actually been proven experimentally. But the problem is that muons (a 200x heavier “relative” of electrons) are unstable, and they are costly (in terms of energy) to create, so it’s difficult to make all this work for practical purposes.
Now Acceleron Fusion, a startup, is working to turn it into a commercial reality. They have raised $24m series A funding, to work on a more efficient process to create muons (a proton accelerator) and increase their lifetime (at higher pressures). Sabine Hossenfelder has just published a video on this approach, and she remains skeptical in the short term, but sounds a bit more positive after discussing the use of another (emerging) technique (laser Wakefield acceleration) to produce the muons. If this could finally work, the implications would (of course) be massive
2 - Demand for AI infrastructure turns Microsoft into the big winner of the 1Q25 “Big Tech” results cycle. The company’s cloud unit had a revenue growth of +20% yoy, in line with consensus, after having missed expectations in 4Q24. Analysts are attributing this to a “big improvement in AI contribution”, which would also justify the large CapEx that the company is reporting ($21.4bn in 1Q25, +50% vs. 1Q24). S Nadella actually explained that the rumors about their cancellation of commitments with data center partners did not mean that they were going to slow down spending
The positive news helped calm investors’ concerns about a potential slowdown in AI infrastructure demand. Before the results presentation, this was the key focus of investors’ discussions, given previous reports that both Microsoft and Amazon might be cutting back on their spending on AI infrastructure, which could signal economic caution and recession fears
Microsoft also announced an ambitious plan to expand their infrastructure in the EU, with the target to operate more than 200 data centers by 2027. In an event in Brussels last Wednesday, Microsoft Vice Chair and President Brad Smith told EU’s policymakers that “AI and cloud data centers represent the next stage of industrialization”, and emphasized that they were already partnering with local companies in the region
3 - However, the debate is still open, with reports of new startups training models without data centers. This week we read an article at Wired describing how a new generation of startups (including Flower AI and Vana) is working to build Large Language Models with no need of centralized AI data centers. They are currently training a 7bn-parameter model using computing capacity spread across hundreds of computers connected over the internet (something that reminds a bit of good old “grid computing”). They also have plans for larger models. Even if this looks pretty ambitious right now, there are clear incentives (including geo-strategic ones) to distribute AI infrastructure a much as possible, so the techniques these guys are developing (e.g. “Photon”, a new tool that makes distributed training more efficient) could be rather useful in the future
4 - More signs that Huawei could be closing the gap (in AI chips) with Nvidia. We’ve already mentioned Huawei’s Ascend 910C chip in previous editions. This week we learned about the company starting to test the next-generation version, the 910D, which is still in early stage, and that would aim to beat Nvidia’s H100 in performance. The H100 was released in 2022, so a significant gap would remain (Nvidia has released the H200 and is now launching the B200, based on the new Blackwell architecture). However, this is one more sign that Huawei is getting closer, and that China could neutralize a significant part of the pressure from US export bans, specially as Chinese AI companies (like DeepSeek) have learned to tweak the software managing the chips to extract additional efficiencies
5 - AI is emerging as the key element in the battle between Meta, Apple and Google (for consumer technology)
Meta’s Andrew Bosworth sees the potential for AI agents (and wearables) to disrupt apps (and smartphones). In a recent interview with VC a16z, Meta’s CTO describes a future in which, rather than getting something done by interacting with an app, users will just describe what they need to an AI agent, and then this agent will do whatever is required to address the need. This is pretty much aligned with visions previously shared (in a B2B context) by people like Salesforce’s Marc Benioff. Even if this obviously would disrupt the current status quo, Bosworth still sees an opportunity for developers, to provide the solutions for the “gaps” that AI agents will find when trying to respond to some user “asks” for which there is no implemented solution yet.
Something similar might have started to impact the Search business model. As users increasingly rely on AI chatbots, rather than search engines, to get information, brands are now trying to find ways to increase their likelihood of surfacing in results from the chatbots. The FT sees this as the “next generation” of “search engine optimization”
Part of the game here could be to signal Apple’s current weaknesses. The comments in the interview happen in the context of a discussion on what will come after the smartphones. And the vision is that these AI agents could lead to different ways for users to interact with technology, or at least to start the interaction (e.g. by talking to a wearable, like Meta’s sunglasses). Apple looks weaker than Meta in this space
Meanwhile, Apple is working to “clean up” its AI activities. We’ve already commented here about the reorganization under way. This week M Gurman gave more details at Bloomberg. Directionally, teams previously centralized within the AI unit are now being segregated and distributed across the company. Apparently, this could help accelerate progress in incorporating AI features into the different AI products (this is what the market expects)
At the same time, Apple might accelerate partnerships with Google (and others) on AI. Everyone thinks there is going to be an integration of Gemini (Google’s AI chatbot) as a built-in option for the iPhone. Google’s CEO S Pichai actually has shared his expectations of this happening “this year”, during court proceedings this week. On a slightly different field, Apple is also partnering with Anthropic (where Google has a significant stake) to adopt the company’s AI-powered coding tools (internally first, but potentially also for third party developers)
6 - New consumer AI apps reveal significant privacy / ethics challenges. This week we learned about at least two of them:
Several startups offering “life-analytics” wearables, powered by AI. J Stern from the WSJ has been trying the services of three different startups (The Bee, Limitless and Plaud) which are already selling wearables that collect everything you say, upload the data to their clouds, analyze this data with AI models, and offer “useful, actionable” insights. Of course, other people’s voices are recorded too, so there are obvious (and massive) privacy-protection implications
Tools for Humanity, a startup backed by Sam Altman, is deploying eyeball-scanning devices across the US. The device, called Orb, is being deployed at selected locations where consumes can verify their identity, and e.g. use the “certificate” for profiles at social media apps, including Tinder, which has been testing a pilot in Japan. The underlying idea is that the need to prove that someone is really human will increase rapidly as AI deepfakes become more convincing (something already happening). However, the fact that users are actually incentivized to scan their eyes, with a payment in Worldcoins (a crypto currency), suggests we should probably prepare for a trade-off of some kind…
7 - Neuralink does a new human implant, on a patient of ALS (a neurological disease that attacks motor neurons in the brain and the spine, often with serious impact on patients’ daily lives, and with an average life expectancy of 2-5 years after diagnosis). The purpose of the trial is to use the Neuralink system to translate the patient’s thoughts into actions, so that he can move a cursor on a computer screen of interact with an iPhone simply by thinking. This is the third time the Neuralink device is implanted in a human
8 - Many startups are working to achieve “fertility on demand”. The focus of most of them is on improving the current, largely manual process of egg freezing + in vitro fecundation, which many women are using, to avoid the clash of having children with their professional careers. For this, several techniques are being developed, from automating the in vitro fecundation process to creating ovarian support cells (enabling in vitro maturation, which is less difficult for patients) or even eggs directly from women’s stem cells (in vitro gametogenesis). Finally, other companies are working on delaying women’s aging and extending their “fertile lives”
9 - A new drug to treat type-1 diabetes has been developed by Australian researchers. A clinical trial with 36 human participants has started in Australia, to test the safety of the drug and the impact of the therapy on a patient's immune system and glucose tolerance. The drug works by calming the auto-immune response that attacks patients’ pancreatic insulin-producing cells
10 - Cyberattacks on two large British retailers, with different degrees of preparedness. Both M&S and Harrods announced they were attacked in the last couple of weeks. This shows how retailers with large customer databases are becoming attractive for hackers, but also how much diversity there can be, within a single industry, in terms of preparedness for these crisis
M&S is struggling to recover after a suspected cyberattack, which has blocked their e-commerce operations (a very substantial segment in the UK). This has had a serious impact on the company’s share price, with a loss of almost £700m in their valuation since the crisis started
Harrods reported signals of a cyberattack last Thursday. The company mentioned that they had detected unauthorized attempts to access some of their systems. Unlike in the case of M&S, the company’s operations were not seriously affected, and both physical stores and e-commerce remained working as normally. This is being seen as a big positive for Harrods, because it would reveal that the company is currently well prepared against cybersecurity crisis
LINKS:
1 - Population & natural resources
Biotech
A new drug to treat type-1 diabetes has been developed by Australian researchers. A clinical trial with 36 human participants has started in Australia. Mum says new diabetes drug would be 'kind of magic' if it works
In the current context of high tariffs, biotech stocks are becoming attractive for investors. Biotech companies like Gilead Sciences have intellectual property registered in the US and also own manufacturing facilities in the country. Even if they pay high taxes for this, these companies are now attractive for investors, as they’re less exposed than others to Trump’s tariffs. Wall Street’s New Tariff Safe Haven: High-Tax Biotech Stocks
Fertility
Many startups are working to achieve “fertility on demand”. The focus of most of them is on improving the current, largely manual process of egg freezing + in vitro fecundation. For this, several techniques are being developed. Fertility on demand
Space
Amazon finally launched the first 27 satellites for Project Kuiper (which aims to compete against SpaceX’s Starlink). As we’ve already reported here, the company’s plan is to have up to 3,200 satellites in orbit, with an investment of more than $10bn… Amazon’s rival to Starlink makes first satellite launch
… So there’s a long way to go, before they can compete with Starlink. Starlink already has 7,000 satellites in orbit, and more than 5m customers, globally. Among the challenges for Kuiper, they need to build the satellites really fast (they have an objective to produce 5 every day). Amazon Wants to Be a Satellite-Internet Powerhouse. It Has a Long Way to Go.
Energy
Energy production is a geo-strategic priority: Hank Paulson, ex-US Treasury Secretary. just published an article at the FT about this, putting energy production as a key geo-strategic tool in the battle against China. He also claims that traditional energy sources will not be enough to address the expected demand growth, so he advocates for “clean” energies (but does not explicitly mention nuclear). Interestingly (see below), China currently dominates the renewable energies’ supply chain, so something else (beyond pushing for solar and wind) would need to be done. Hank Paulson: Clean energy will be critical to winning the AI race with China
Nuclear
Could muons make “cold” nuclear fusion possible? Substituting electrons with muons to orbit hydrogen or helium atoms’ nuclei makes it easier to force these nuclei to “merge” and trigger a nuclear fusion reaction, even at room temperatures (with no need to create and confine ultra-high temperature plasma). This has actually been proven experimentally, and Now Acceleron Fusion, a startup, is working to turn it into a commercial reality. Sabine Hossenfelder (@skdh) on X
Renewables
For now, China dominates the “renewable energy” supply chain. The country’s exports of solar panel and other “green energy” technologies to emerging markets are growing fast. And the expectation is that they will grow even faster, with the current US tariffs effectively neutralizing potential sales to the Americans. China’s clean tech exports to emerging markets surged in 2024, data shows
South Korean rivals to Chinese suppliers expect to benefit from current tariffs in the US. An example are batteries for energy storage in electric grids, which are key for backbone networks to cope with fluctuating wind and solar energy supplies. China dominates the market for this technology (90% global market share, 80% in the US, 75% in Europe), but now tariffs create an opportunity (at least in the US) that South Korean want to capture. China and South Korea extend battery battle from EVs to grid storage
Vestas, Europe’s largest turbine maker, asks for an acceleration of the “green transition”. They claim that Europe is not moving fast enough to deliver on the “net zero” target for 2050. Permits / regulation / bureaucracy are mentioned as one of the problems. However, after events like the power outage in Spain and Portugal this week, it is not clear that it will be easy to relax regulatory pressure… Vestas warns wind industry is falling behind global climate goals
2 - Efficiency & Productivity
New Transport Technologies
Autonomous Cars
Toyota partners with Waymo to explore the use of self-driving technology in personally owned vehicles. Apparently Toyota would be interested in “enhancing” the functionalities of its cars with self-driving features. They could also become a provider for Waymo’s fleets of robotaxis. Waymo, Toyota strike partnership to bring self-driving tech to personal vehicles
In China, Pony AI is making robotaxis profitable. Pony AI, one of the three key companies in China’s robotaxis industry, has managed to cut the cost per unit of its autonomous cars from approx. $140,000 to $40,000. They now claim to be in a position to reach profitability. Chinese Robotaxi Firm Pony AI on Faster Track to Profit After Cost Cuts, CTO Says
Aurora Innovations shows that the revolution might start with autonomous cargo trucks. The company has just began operations with 3 round-trips in the route between Houston and Dallas (around 200 miles). Many people see cargo trunks as a first step before expanding into passenger traffic and personal cars. AI-Driven 18-Wheelers in Texas Have No Room for Error
Computing Infrastructure
IBM will spend $30bn in the next five years to build mainframes and quantum computers. This is part of an overall $150bn that the company committed to spend in the US, in what might be seen as a signal for the Trump administration. This is interesting, as it points to the two extremes of the computing technology spectrum. Actually, the money for quantum computing will probably be fully dedicated to R&D, given the timelines. IBM to Spend $150 Billion in U.S. Over Next Five Years
Data Centers
Demand for AI infrastructure turns Microsoft into the big winner of the 1Q25 “Big Tech” results cycle. The company’s cloud unit had a revenue growth of +20% yoy, in line with consensus, after having missed expectations in 4Q24. Microsoft shares jump after strong AI demand lifts cloud unit
The positive news helped calm investors’ concerns about a potential slowdown in AI infrastructure demand. Before the results presentation, this was the key focus of investors’ discussions, given previous reports that both Microsoft and Amazon might be cutting back on their spending on AI infrastructure. Microsoft and Amazon Capex in Focus Amid Potential AI Pullback
Microsoft also announced an ambitious plan to expand their infrastructure in the EU, with the target to operate more than 200 data centers by 2027. In the announcement, Microsoft Vice Chair and President Brad Smith said that “AI and cloud data centers represent the next stage of industrialization”. Microsoft Touts EU Data Center Expansion, Pledges to Protect European Operations
However, the debate is still open, with reports of new startups training models without data centers. This week we read an article at Wired describing how a new generation of startups (including Flower AI and Vana) is working to build Large Language Models with no need of centralized AI data centers. These Startups Are Building Advanced AI Models Without Data Centers
Chips
After the impact of export bans on its sales in China, Nvidia is preparing for more blows. The US government’s “AI diffusion” rules, which try to avoid sales of strategic technologies to China via third countries, will come into effect in less than 1 month, restricting the company’s ability to sell in several “US-friendly” markets. This is expected to be a second blow for Nvidia, after the impact of previous sanctions on their direct sales to China. Nvidia’s Trump Worries Go Far Beyond China
Analysts are starting to get nervous. As an example, Nvidia just received a “Sell” rating from Seaport Global Securities, a broker, supported by the thesis that “AI has been priced in for now”. Nvidia Gets Rare Sell Rating as Seaport Says AI Fully Priced In
Artificial Intelligence
Everyone seems to accept that the partnership between Microsoft and OpenAI is now strained. There are disagreements on the key elements of the alliance, with OpenAI looking for alternative providers of computing infrastructure, and setting restrictions on Microsoft to use its newest models. Altman and Nadella, Who Ignited the Modern AI Boom Together, Are Drifting Apart
AI: Apps, Agents
B2C
New consumer AI apps reveal significant privacy / ethics challenges. This week we learned about at least two of them:
Several startups offering “life-analytics” wearables, powered by AI. J Stern from the WSJ has been trying the services of three different startups (The Bee, Limitless and Plaud) that sell wearables that collect everything you say, upload the data to their clouds, analyze this data with AI models, and offer “useful, actionable” insights. I Recorded Everything I Said for Three Months. AI Has Replaced My Memory.
Tools for Humanity, a startup backed by Sam Altman, is deploying eyeball-scanning devices across the US. The device, called Orb, is being deployed at selected locations where consumes can verify their identity, and e.g. use the “certificate” for profiles at social media apps, including Tinder, which has been testing a pilot in Japan. Altman-Backed Startup Rolls Out Eyeball-Scanning Tech Across US Sam Altman’s eyeball-scanning project World makes US debut
AI is emerging as the key element in the battle between Meta, Apple and Google (for consumer technology)
Meta’s Andrew Bosworth sees the potential for AI agents (and wearables) to disrupt apps (and smartphones). In a recent interview with VC a16z, Meta’s CTO describes a future in which, rather than getting something done by interacting with an app, users will just describe what they need to an AI agent, and then this agent will do whatever is required to address the need. What Comes After Mobile? Meta’s Andrew Bosworth on AI and Consumer Tech
Something similar might have started to impact the Search business model, as users increasingly rely on AI chatbots, rather than search engines. Brands target AI chatbots as users switch from Google search
Meanwhile, Apple is working to “clean up” its AI activities. Directionally, teams previously centralized within the AI unit are now being segregated and distributed across the company. Apple Begins Breaking Up Its AI Team With Robotics, Siri Changes
At the same time, Apple might accelerate partnerships with Google (and others) on AI. Everyone thinks there is going to be an integration of Gemini (Google’s AI chatbot) as a built-in option for the iPhone. Google’s CEO S Pichai actually has shared his expectations of this happening “this year”, during court proceedings this week. Google Eyes Gemini-iPhone AI Deal This Year, Pichai Tells Court
On a slightly different field, Apple is also partnering with Anthropic (where Google has a significant stake) to adopt the company’s AI-powered coding tools (internally first, but potentially also for third party developers). Apple, Anthropic Team Up to Build AI-Powered ‘Vibe-Coding’ Platform
Finally, Meta has just launch its stand-alone AI app, which now integrates the “Meta View” functionality that controls the Ray-Ban smart glasses, and is seen as a vehicle for Meta to get more AI users, beyond the ones captured within WhatsApp or Instagram. Meta launches stand-alone AI app to take on ChatGPT
In China, the model to integrate AI into consumer apps is not vertically integrated. According to these comments at the FT, this would have lessons for investors, mostly in the direction to reject the current valuations of Elon Musk’s X (the app / social network) and xAI (the company building the AI model -Grok). How China has changed the game for AI valuations
B2B / Agents
AI agents could have already started to replace humans. According to this FT article, something like this would be behind the positive results that Big Tech companies working on AI have published in the last few days. Your pushy AI intern is ready for a promotion
The next step would be creating corporate processes to enable multi-agent collaboration. Steven Rosenbush at the WSJ claims that individual agents have already been deployed (or used by employees) at many companies. However, to make the most of agents, companies would need to set up processes enabling collaboration among multiple agents. Part of what’s needed (according to the article) involves eliminating humans from the workflows, so there will probably be trade-offs and potential risks that companies should take care of, too. AI Agents Are Learning How to Collaborate. Companies Need to Work With Them
Microsoft claims to be building 30% of the company’s internal code with AI agents. This was revealed by S Nadella in a fireside chat with M Zuckerberg this week. Previously, the company’s CTO had commented that he expected 95% of the company’s code to be generated by AI systems by 2030. Microsoft CEO says up to 30% of the company's code was written by AI
Finance is another field where AI is penetrating fast:
Norway’s Sovereign Fund ($1.8tn in managed assets) expects to save a fifth of its annual trading costs by using AI to trade less frequently and more efficiently. The fund is one of the biggest equity investors in the world, owning on average 1.5 per cent of every listed company globally. Norway’s oil fund aims to save $400mn of trading costs using AI
Rogo, an AI startup, is commercializing a chatbot that aims to fully replace an investment banker. They just raised $50m in new capital, at a valuation of $350m. The tool has already been deployed at investment banks Moelis and Nomura, and is one more reminder of the emerging dilemma about “junior jobs” disappearing at these companies (which closes the traditional path for starting a professional career there). Meet your new investment banker: an AI chatbot
Physical AI: Robots
Military drones are being built, which will be able to autonomously pick and hit targets. Stark, a German startup, is building autonomous military drones, and expects them to be prepared for deployment in real-world operations soon. Many unmanned systems can already navigate autonomously and track moving targets, but the new versions being developed will also be able to pick and hit targets without human supervision. So there are plenty of ethical implications behind this… Fully autonomous strike drones within technological reach, says German start-up
AI: Foundational Models
Meta is launching a commercial AI computing service, based on an API giving access to the company’s Llama models. For now, Meta is focused on use cases requiring fast response / inference speeds, as opposed to slower “reasoning” models currently being promoted by OpenAI or Google. This is partly due to a weakness of Meta (their “reasoning” model is still under development), but the results (18x faster inference speeds) are impressive in any case. Meta unleashes Llama API running 18x faster than OpenAI: Cerebras partnership delivers 2,600 tokens per second
AI: Security & Safety
Cyberattacks on two large British retailers, with different degrees of preparedness:
M&S is struggling to recover after a suspected cyberattack, which has blocked their e-commerce operations (a very substantial segment in the UK). M&S cyber crisis wipes almost £700mn off retailer’s valuation
Harrods reported signals of a cyberattack last Thursday. The company mentioned that they had detected unauthorized attempts to access some of their systems. Unlike in the case of M&S, the company’s operations were not seriously affected. Harrods is latest UK retailer to suffer cyber attack
Intelligence Augmentation
Brain-Computer Interfaces
Neuralink does a third human implant, on a patient of ALS (a neurological disease that attacks motor neurons in the brain and the spine). The purpose is to use the Neuralink system to translate the patient’s thoughts into actions, so that he can move a cursor on a computer screen of interact with an iPhone simply by thinking. Exclusive: Mr. Smith Gets A Neuralink Implant
The guy has even started to tweet after receiving the implant. Bradford G Smith (Brad) (@ALScyborg) on X
Whole Brain Emulation
Understanding mice’s brains, as a first step to build representations of human ones. In an interview this week, a researcher working to build a digital representation of the human brain explains the project and how it is related to a recently published result about how the visual cortex of mice works. A Brain Researcher's Bid to Make Digital Twins of Humans
3 - Economic / Business trends
Tech & Geopolitics
More signs that Huawei could be closing the gap (in AI chips) with Nvidia.
The company has started to work in their next-gen AI chip. We’ve already mentioned Huawei’s Ascend 910C chip in previous editions. This week we learned about the company starting to test the next-generation version, the 910D, which is still in early stage, and that would aim to beat Nvidia’s H100 in performance. Nvidia’s H200 and the newer B200 are still ahead, but the gap seems to be closing. Exclusive | China’s Huawei Develops New AI Chip, Seeking to Match Nvidia
Meanwhile, they’ve started to sell CloudMatrix 384, a “cluster” aggregating 384 910C chips. This is the platform we discussed recently, which chip guru Dylan Patel views as “an AI system capable of beating Nvidia’s” (yes, at the expense of higher energy consumption, but still…) Huawei delivers advanced AI chip ‘cluster’ to Chinese clients cut off from Nvidia
Trump would be considering to simplify US chip export bans. The current rule, from Biden’s times, classifies countries into three tiers and sets constraints depending on each tier. Trump wants to replace this with a single global licensing regime, complemented with 1-to-1 agreements with the different countries. Exclusive: Trump officials eye changes to Biden's AI chip export rule, sources say
Apple confirmed that iPhones to be sold in the US won’t be built in China. But they won't be manufactured in the US, either. The beneficiaries will be India and Vietnam. Apple Says Most of Its Devices Shipped Into U.S. Will Be From India, Vietnam
Emerging economic models
Nvidia CEO shared his vision of a world where “AI Factories” will play a critical role. According to this, societies will need “AI model factories” to build the software components (advanced, possibly specialized AI models) that will power all kinds of products (from cars to consumer electronics). He also expects that building these factories will be a source of new jobs. Nvidia CEO Says All Companies Will Need ‘AI Factories,’ Touts Creation of American Jobs
A couple of lawsuits against Meta show us a picture of what is coming:
The lab behind the model is sued for something the model said. A political activist in the US has sued Meta alleging that Meta AI linked him to the Jan 6th 2021 riots in Washington. Exclusive | Activist Robby Starbuck Sues Meta Over AI Answers About Him
The lab is sued for using copyrighted material to train its models. A dozen authors are suing Meta for having used LibGen, a “shadow library” containing millions of books, to train the Llama models. Meta lawsuit poses first big test of AI copyright battle