Headlines this week - Apr 13, 2025
A look at how capital is being deployed across future opportunities
This week in the future:
1 - Nuclear power is back, driven by AI demand, and redesigned with AI
Although there is still debate on this, the expected energy demand in data centers remains huge. In a new report, the International Energy Agency predicts that, by 2030, the data centers powering AI will require more energy than Japan consumes today. Only in the US, data centers are expected to account for more than 50% of the energy demand growth through 2030
Nuclear energy is a straightforward way to address these needs. Indeed, the technology industry has become a leading advocate for nuclear energy. Microsoft and Amazon are working to re-open old nuclear plants, OpenAI’s Sam Altman is supporting a nuclear energy startup (Oklo) that is building a “small modular reactor”, and even Bill Gates is the co-founder of another startup, TerraPower, looking to create smaller, cheaper and safer reactors
Even environmentalists may be reconsidering their previous views. A recently published book, reviewed by The New Yorker this week, describes how several groups of “evangelists” in the US are spreading the message that the fear of radiation has actually been more dangerous than nuclear technology itself. Yes, some of these groups are subsidized by energy companies, but it also looks like factual data is making some people change their minds
AI is emerging as a tool to design new plants and run the existing ones. The US Energy Department’s Argonne National Laboratory has developed an AI-based tool that can assist with reactor design and help operators run nuclear plants. This is expected to help both with projects to build new nuclear facilities and with initiatives to “resurrect” or extend the lifetime of the current American plants, which have an average age of 42 years, and badly require some improvements
2 - Stanford University’s Institute for Human-Centered Artificial Intelligence published “The state of AI in 2025”
Some insights from the report:
(1) The US is still the global leader by number of LLMs (40, vs. 15 in China and 3 in Europe);
(2) Training costs keep rising, driven by the scale up in parameter count, training time, and amount of training data -but the analysis does not (yet) include DeepSeek;
(3) Inference costs are going down, and this opens the door to implement more reasoning (that takes advantage of these cost reductions) in the models;
(4) AI’s energy costs (and carbon footprint) are also rising, in spite of new energy efficiencies;
(5) The performance gap between the US and China is reducing (from 9.3% to 1.7% in just 1 year);
(6) We are struggling to find appropriate test to measure the models’ performance, but the latest effort (“Humanity’s Last Exam”) is working relatively well for now;
(7) Websites have started to restrict access to their data sets, and this could force AI companies to use alternative approaches (vs. “more data”) to train the models;
(8) Investors are pouring huge amounts of money on AI: private investment on AI reached a $150.8bn peak in 2024;
(9) Companies have not yet achieved big returns out of their AI investments -a recent McKinsey survey shows that most had savings of less than 10%;
(10) AI models (alone) are diagnosing better than human doctors and models combined), and this may be slowing down adoption in medical contexts
3 - A debate on the coming “AI revolution”. Will “super intelligence” be created? Will it replace humans in the “physical world”?
People are still debating the “AI2027” report we commented here last week. The document explores scenarios of AI development in the coming months, and reaches the conclusion that the development of a “super intelligence” (better than humans at almost anything) by 2027 would be very likely
Even if “super intelligence” is achieved, it is not clear that it will be able to overcome the barriers for impact in the physical world. Spanish blogger Alberto Romero has published a nice analysis in which he shares his skepticism on the timelines (2027) that the report proposes. According to him, “there is unassailable friction between super intelligence and revolution. When you solve intelligence, the bottleneck for effective change doesn't disappear; it moves somewhere else”. Among the new bottlenecks: politics, social resistance, cultural barriers, logistics (in the world of atoms) or bureaucracy. So he concludes that the revolution, if it happens, will take way more time…
4 - In any case, it is increasingly clear that the future of AI will depend on what happens in China
The “AI2027” report might have underestimated China’s capabilities. This is another point raised by Alberto Romero in his post. He claims that “China has likely already caught up on the algorithmic front” (see the message about the gap narrowing down in the Stanford report, discussed above). And this could have implications on many things, including the safety issues addressed by the “AI2027” authors
We keep receiving confirmations about China’s strength in AI models. This week we have news about:
DeepSeek’s research to further reduce the costs of training the models, in collaboration with Tsinghua University (probably the top one for technology in China). The methodology is based on Reinforcement Learning with rewards for more accurate and understandable responses
ByteDance’s plan to leverage their access to a massive amount of consumer data to become an AI leader. ByteDance’s apps have more than 1bn users globally (170m in the US), and provide a permanent source of data that can be exploited to build better models. They would be in a similar position to Meta’s (yes, maybe quite smaller, but also with data from 700m Chinese users -something that Meta does not have). ByteDance is also spending billions in data centers
5 - Brain research continues making progress. This could lead to more powerful / efficient AIs and to better connections between human intelligence and AIs
It is becoming easier / cheaper to map human brain’s network connections. This week a team from the Baylor College of Medicine published a paper in Nature describing a software package that they have developed and that simplifies the process of “reading” the wiring of the human brain. This is part of the efforts to fully map the brain’s neural network, and build systems capable to emulate a whole human brain. Understanding how the brain works could open the door to stronger, more efficient AI systems
Brain-computer interfaces keep moving ahead to become commercially viable. As an example this week, a team at Georgia Tech has developed something close to what we could describe as a “wearable brain-computer interface”: an almost imperceptible sensor that can be inserted into the minuscule spaces between hair follicles and slightly under the skin
Therapeutic applications are the “quick wins”. As we’ve already discussed here, therapeutic solutions for brain-related illnesses are the short-term objective of a lot of research teams in the field. This week we learned about a team in Stanford that has been able to re-create a human neural circuit in a lab dish. Of course, this could be related with future “neurocomputing” chips. But first it could have more immediate applications e.g. understanding how pain is transmitted to the brain, which can be used to develop new therapeutics for chronic pain
6 - AI-based military systems are becoming a key area of focus for AI companies. This could be the actual (unconfessed) driver of the current “AI race” between the US and China (with Europe far behind for now)
OpenAI does not discard working with the Pentagon. Last Thursday, at the Vanderbilt Summit on Modern Conflict and Emerging Threats, Sam Altman said he (and his company, OpenAI) would not rule out helping the Pentagon develop a new weapons platform. Yes, he also said that he thought no one wanted AI systems to make weapons’ decisions, but he argued that the world could get “really weird” to justify why he would “never say never”
Generative AI is already being used for intelligence operations. A division of the Marines, during a recent expedition throughout the Pacific, used a model produced by the defense-tech company Vannevar Labs to screen thousands of pieces of open-source intelligence—nonclassified articles, reports, images, videos—collected in various countries. The company claims to be collecting terabytes of data every day in 80 different languages and across 180 countries. And they’re a Pentagon contractor
7 - Robots are a central part of this “AI race”, and investors are excited
Figure AI, an “autonomous robot” startup with no revenues, is raising capital at a valuation of $40bn. Their pitch is that they will be able to build 200,000 autonomous robots by 2029, and their annual revenues will be around $9bn by then. They have already signed a contract with BMW. For now, they have released some impressive videos, but not much more. They already raised $675m at a valuation of $2.5bn last year (so if this goes as they expect, it would mean a 16x increase in value, in just one year)
Investment banks are excited about the opportunity: Goldman Sachs expects the global humanoid robot market to be worth $205bn by 2035. Bernstein estimates annual robot sales of up to 50m in 2050. Citibank forecasts 648m humanoid robots by 2040, and Bank of America (in an awesome display of long-term vision) predicts 3bn by 2060. They’re mainly looking at Chinese companies (see below) and a few American ones (including Tesla, Meta, and startups like Figure AI (see previous point), Boston Dynamics or Agility Robotics
China may have an advantage in robots. Leveraging their superiority in all things linked to “physical manufacturing”, the Chinese would be more advanced than the Americans in humanoid robots. Companies like Unitree, AgiBot, Engine AI and UBTech are the leaders. There might be strong synergies between robots and electric vehicles, so China would be well positioned to dominate both fields
8 - There are also some signs that the “AI race” could lead to AI safety issues. This was one of the concerns of the authors of the “AI2027” scenarios. And it seems on the way to become a real threat:
OpenAI is reducing the time dedicated to testing its models, before releasing them. The reason for this would be the need to accelerate progress, as Chinese models are catching up and there would be a risk that the Chinese would reach “super intelligence” (a threshold from which models would be able to self-improve) before. But the risks are obvious.
(At a very different level), in Europe, the EU wants to reduce the regulatory burden on AI startups. Apparently concerned about the trade-offs between regulation and innovation, the European Commission (in an internal document) is discussing the “opportunity to minimize the potential compliance burden of the AI Act, particularly for smaller innovators”. In principle, this would only be related to (reducing) the bureaucracy that companies have to deal with, in relation to the Act, so safety would not be compromised…
Meanwhile, new safety / security / privacy risks keep emerging
AI models can be used to create more powerful threats. As an example, OpenAI is already being used to create more effective spam, through text messages that are able to bypass current spam filters
Geopolitical issues are increasing risks. E.g. China has implicitly admitted they’re behind a recent series of cybersecurity attacks to US infrastructures. This happened in a bilateral meeting last December, in which the Chinese delegation linked the attacks to the “escalating US support” for Taiwan
Waymo intends to use internal cameras in the cars to collect passengers’ data to train their models. At least according to a researcher who claims to have found a yet unreleased version of Waymo’s privacy policy. So passengers hiring an autonomous Waymo car might have to expect personalized data related to how they behaved in the car. This all sounds a bit creepy…
9 - An opportunity is emerging for pharma companies to partner with (or acquire) biotech startups
Current uncertainties are making it difficult to think about IPOs in biotech. Biotech startups are affected both by volatility in the financial markets and uncertainty in the political side (worse after R F Kennedy was appointed by the Trump administration), with budget cuts now on the way at the National Institutes of Health. All this creates a difficult scenario for potential IPOs. At the same time, pharma companies badly need innovations (e.g. like those triggered by the use of AI to develop new medicines). So the atmosphere is appropriate for partnerships, and maybe acquisitions too
Investors could favor these deals. They are currently nervous about biotech startups’ business models, combining risky science, costly funding, political uncertainty and long waits for returns
The threat of China dominating the industry could be a catalyst. Last Tuesday, a bipartisan commission at the US Congress called for $15 billion in funding for biotech research and manufacturing over the next five years, justified by the need to avoid paying “a very heavy price in terms of economic and national security”, given how China is progressing. So there are political incentives to help Biotech startups succeed
10 - The debate on “freedom cities” is back. Now with focus on Greenland
Silicon Valley investors are proposing to build a libertarian 'freedom city' in Greenland. Discussions would be in early stages. Names like Marc Andreessen and Peter Thiel are being mentioned as potential supporters for the initiative. And the plan could include a regulation-free hub for the development of artificial intelligence, autonomous vehicles, space launches, micro nuclear reactors and high-speed rail
LINKS:
1 - Population & natural resources
Biotech
An opportunity is emerging for pharma companies to partner with (or acquire) biotech startups
Current uncertainties are making it difficult to think about IPOs in biotech. At the same time, pharma companies badly need innovations (e.g. like those triggered by the use of AI to develop new medicines). Market Volatility to Drive Closer Collaboration Between Biotech Startups, Large Pharma
Investors could favor these deals. They are currently nervous about biotech startups’ business models, combining risky science, costly funding, political uncertainty and long waits for returns. How to Play the Biotech Meltdown in the Age of RFK Jr. and Tariffs
The threat of China dominating the industry could be a catalyst. Last Tuesday, a bipartisan commission at the US Congress called for $15 billion in funding for biotech research and manufacturing over the next five years. China’s Biotech Advances Threaten U.S. Dominance, Warns Congressional Report
Healthcare
New wearables on the way to monitor people’s health. A team at Northwestern University (USA) has developed a device that monitors gases interacting with the skin, which is expected to help with understanding skin health and aiding patients with various conditions. Revolutionary Wearable Device Tracks Your Skin’s Vital Signs
Space
Amazon Project Kuiper’s battle with Starlink will have to wait. The launch of the company’s first satellites was cancelled last Wednesday due to bad weather. They are now reviewing the next opportunity for a launch. Launch of First Amazon Project Kuiper Internet Satellites Is Scrubbed
2 - Efficiency & Productivity
Energy
Although there is still debate on this, the expected energy demand in data centers remains huge. In a new report, the International Energy Agency predicts that, by 2030, the data centers powering AI will require more energy than Japan consumes today. AI Boom to Fuel Surge in Data Center Energy Needs, IEA Says
Nuclear
Nuclear energy is a straightforward way to address these needs. Indeed, the technology industry (Microsoft, Amazon, Sam Altman, Bill Gates) has become a leading advocate (and financial supporter) for nuclear energy.
Even environmentalists may be reconsidering their previous views. A recently published book, reviewed by The New Yorker this week, describes how several groups of “evangelists” in the US are spreading the message that the fear of radiation has actually been more dangerous than nuclear technology itself. Environmentalists Are Rethinking Nuclear. Should They?
AI is emerging as a tool to design new plants and run the existing ones. The US Energy Department’s Argonne National Laboratory has developed an AI-based tool that can assist with reactor design and help operators run nuclear plants. Nuclear Power Is Back. And This Time, AI Can Help Manage the Reactors.
Renewables
Renewable energies are already providing almost one third of global electricity. The number was published this week by Ember (an energy think tank). Renewables provided record 32% of global electricity in 2024, Ember says
New Transport Technologies
Electric Vehicles
A fierce competition in the Chinese local markets is helping Chinese EV companies (like BYD) dominate the world. The pressure of competition in China is catalyzing a wave of vehicle innovations (including self-driving features and ultra-fast charging) that position these companies in advantage to dominate Western markets. The relentless innovation fuelling China’s ‘brutal’ car wars
Autonomous Cars
Traditional brands may have a hope, in partnering with software startups. As an example this week, Wayve, a British startup developing software to control autonomous cars, is partnering with Nissan to deploy autonomous cars, starting in 2027. UK AI start-up Wayve clinches first carmaker deal with Nissan
Artificial Intelligence
Stanford University’s Institute for Human-Centered Artificial Intelligence just published “The state of AI in 2025”. This article at IEEE Spectrum neatly summarizes the main conclusions, with 12 interesting charts. These 12 Eye-Opening Graphs Reveal the State of AI in 2025
People are still debating the “AI2027” report we commented here last week. Spanish blogger Alberto Romero published a nice analysis in which he shares his skepticism on the timelines that the report proposes. Even God Can’t Skip the Bureaucrats
Profits erosion from tariffs affecting Big Tech companies could have an impact on AI investment. E.g. both advertising and e-commerce revenues (key revenue engines for Big Tech firms) could be affected by tariffs. Tariffs Are Still Coming for the AI Boom
AI: Apps
B2C
AI has reached the mass market: The ChatGPT app is now more downloaded than Instagram and TikTok. ChatGPT became the most downloaded app globally in March
OpenAI might still be thinking about a consumer device. They have been considering to acquire io Products, a hardware design startup run by ex-Apple Jony Ive, for $500m. OpenAI been collaborating with io Products in designs for future AI-powered consumer devices. Sam Altman’s OpenAI mulls buying ex-Apple design guru Jony Ive’s AI device startup for $500M: report
AI helps merge ads with user-generated content. Snap just released “sponsored AI lenses”, a tool that makes it possible for users to create personal content linked to specific brands. This puts advertising in collision course with user-generated content, and could be a sign of what is to come… Snap’s New Generative AI Ad Format Emphasizes Engagement Over Scale
Netflix is powering content search with AI. With this you will be able to look for shows using more specific terms, including the way you feel… Netflix Tests New AI Search Engine to Recommend Shows, Movies
Website building is now democratized. WordPress just launched a GenAI tool for anyone to build a website easily. WordPress.com will now build you a full website with AI
B2B
AI-based military systems are becoming a key area of focus for AI companies
OpenAI does not discard working with the Pentagon. Last Thursday, at the Vanderbilt Summit on Modern Conflict and Emerging Threats, Sam Altman said so. OpenAI’s Altman Won’t Rule Out Helping Pentagon on AI Weapons
Generative AI is already being used for intelligence operations. E.g.: The US Marines are using a model by the defense-tech company Vannevar Labs to screen thousands of pieces of open-source intelligence. Generative AI is learning to spy for the US military
Andy Jassy (Amazon’s CEO) believes that only companies using GenAI to “reinvent” the customer experience will remain competitive. He said so in his annual letter to Amazon shareholders. Amazon CEO Says Generative AI Will Reinvent Every Customer Experience
All kinds of content creation are being impacted by AI, right now:
News production (and reading). As this New Yorker article suggests, there may be value in integrating AI systems in the process, but only if the industry proceeds with care (there are many risks emerging, in the transition). Will A.I. Save the News?
Music creation. YouTube is deploying an AI tool for creators to generate instrumental music to back their videos. This looks like a first step towards a future in which users will be able to add their own music too. YouTube rolls out a free AI music-making tool for creators
Literature / writing. Yes, this article only mentions romance novels, but why should it stop there? The Romance Publisher Dreaming of an AI-Driven Dynasty
AI coding tools keep improving. E.g. Google’s Gemini Code Assist, which will now have “agentic” capabilities. The tool will now be able to deploy AI “agents” that can take multiple steps to accomplish complex programming tasks. Gemini Code Assist, Google's AI coding assistant, gets 'agentic' abilities
AI is affecting companies’ recruiting. Shopify has said they will not hire unless the job can’t be done by an AI. Shopify Says No New Hires Unless AI Can’t Do the Job
AI: Robots
Robots are a central part of this “AI race”, and investors are excited
Figure AI, an “autonomous robot” startup with no revenues, is raising capital at a valuation of $40bn. Their pitch is that they will be able to build 200,000 autonomous robots by 2029, and their annual revenues will be around $9bn by then. All this euphoria could be explained by the vision (common among investment banks right now) that robots could become “the next smartphones”. The $40 Billion Startup Mystery Shaking Up Silicon Valley
China may have an advantage in robots. Leveraging their superiority in all things linked to “physical manufacturing”, the Chinese would be more advanced than the Americans in humanoid robots. China gains dexterous upper hand in humanoid robot tussle with US
Samsung is partnering with Google to build a home robot, the Ballie. The product is a soccer-ball-shaped, AI-powered home robot that can project video onto walls. Samsung Taps Google AI to Launch Long-Promised Ballie Robot With Video Projector
AI: Foundational Models
We keep receiving confirmations about China’s strength in AI models:
The performance gap between the US and China in AI models is reducing (from 9.3% to 1.7% in just 1 year). The AI Race Has Gotten Crowded—and China Is Closing In on the US
DeepSeek is working to further reduce the costs of training the models, in collaboration with Tsinghua University. DeepSeek and Tsinghua Developing Self-Improving AI Models
ByteDance (TikTok’s owner) wants to leverage its access to a massive amount of consumer data, to become an AI leader. ByteDance’s apps have more than 1bn users globally (170m in the US). How TikTok’s Parent, ByteDance, Became an A.I. Powerhouse
OpenAI is improving the models’ memory. This is a space where also Google is innovating. Companies expect these features to radically improve consumer experience. OpenAI updates ChatGPT to reference your past chats
Open Source models could soon lead the industry. Deep Cogito, a new AI startup in the US, has launched a new line of open source models, based on Meta’s Llama 3.2, and including reasoning capabilities, which are already topping the model performance charts. New open source AI company Deep Cogito releases first models and they’re already topping the charts
AI: Security & Safety
Could the “AI race” be creating AI safety issues? This was one of the concerns of the authors of the “AI2027” scenarios. And it could be on the way to become a real threat:
OpenAI is reducing the time dedicated to testing its models, before releasing them. The reason for this would be the need to accelerate progress, as Chinese models are catching up. OpenAI slashes AI model safety testing time
(At a very different level), in Europe, the EU wants to reduce the regulatory burden on AI startups. Apparently concerned about potential trade-offs between regulation / bureaucracy and innovation / speed. Europe wants to lighten AI compliance burden for startups
Meanwhile, new safety / security / privacy risks keep emerging
AI models can be used to create more powerful threats. As an example, OpenAI is already being used to create more effective spam (able to bypass current filters). OpenAI helps spammers plaster 80,000 sites with messages that bypassed filters
Geopolitical issues are increasing risks. E.g. China has implicitly admitted they’re behind a recent series of cybersecurity attacks to US infrastructures. Exclusive | In Secret Meeting, China Acknowledged Role in U.S. Infrastructure Hacks
Waymo intends to use internal cameras in the cars to collect passengers’ data to train their models. At least according to a researcher who claims to have found a yet unreleased version of the autonomous car company’s privacy policy. Waymo may use interior camera data to train generative AI models, sell ads
A “social backlash” could be coming against social media companies, with users increasingly concerned about their apps’ negative effects on people (including young kids). This week Parmy Olson compared this to the previous move against the tobacco industry… Social Media's 'Big Tobacco Moment’ Is Coming
AI: Infrastructure
Companies and governments remain committed to massive investments in AI infrastructure, in spite of the uncertainties created by DeepSeek
Alphabet confirmed their CapEx guidance of $75bn for this year. Some observers were expecting a change, in the current context of Trump’s tariffs. Alphabet reaffirms $75 billion spending plan in 2025 despite tariff turmoil
A reason for this is that they need to fight the impression that they could lose the “AI race”. E.g. some investors believe AI chatbots could be on the way to disrupt search, and this is penalizing Google’s stock right now. Alphabet Can’t Shake Off AI Concerns Even With Low Multiple
Meanwhile, the EU views AI infrastructure as a key component for technology self-sufficiency (a priority in Europe these days). EU Bets on Gigafactories to Catch Up With U.S., China in AI Race
AI: Chips
Tariffs are a key concern for the chip industry at this point in time, with all kinds of strange / unpredictable effects
China is trying to create incentives for US chipmakers to manufacture out of the US. They just announced that the tariffs to sell in China will depend on where the chip is manufactured (not on the country where the vendor is based). US chipmakers outsourcing manufacturing will escape China's tariffs
Samsung is profiting from customers stockpiling chips in anticipation of tariff execution. Samsung boosts profits as customers stockpile chips ahead of tariffs
The perspective of tariffs blurred an otherwise brilliant set of results for TSMC in 1Q2025, with revenues beating consensus. Actually, this was just an advance of the full results to be published on Apr 17. But the stock market is now focused on following tariff announcements, and results are not affecting it so much… TSMC's first-quarter revenue surges 42%, slightly ahead of forecasts
Intelligence Augmentation
Brain-Computer Interfaces
Brain-computer interfaces keep moving ahead to become commercially viable. As an example this week, a team at Georgia Tech has developed something close to what we could describe as a “wearable brain-computer interface”. New Wearable Brain-Computer Interface | Research
Whole Brain Emulation
It is becoming easier / cheaper to map human brain’s network connections. This week a team from the Baylor College of Medicine published a paper in Nature describing a software package that they have developed and that simplifies the process of “reading” the wiring of the human brain. NEURD - Proofreading the map of the brain
Therapeutic applications are the “quick wins”. A team in Stanford has been able to re-create a human neural circuit in a lab dish. This could have help understanding how pain is transmitted to the brain, which could be useful to develop new therapeutics for chronic pain. Scientists recreate brain circuit in lab for first time
3 - Economic / Business trends
Tech & Geopolitics
Is TikTok US part of the China-US negotiations for tariffs? The new chapter in the fight to separate TikTok US from its Chinese owner is seen in this New Yorker article as part of the currently escalating trade war. TikTok would be a bargaining chip in the negotiations. And commercial interests, rather than consumer safety, is what would really be at stake. TikTok and the Retreat from Technological Globalization
Emerging economic models
The debate on “freedom cities” is back. Now with focus on Greenland. Silicon Valley investors are proposing to build a libertarian 'freedom city' in Greenland, with a plan to build a regulation-free hub for the development of artificial intelligence, autonomous vehicles, space launches, micro nuclear reactors and high-speed rail. Exclusive: Greenland ‘Freedom City’? Rich donors push Trump for a tech hub up north