Headlines this week - Mar 9, 2025
A look at how capital is being deployed across future opportunities
This week in the future:
The battle over chips for “technology self-sufficiency” is heating up:
This week TSMC announced a $100bn investment in the US, to build local chip production, packaging and R&D facilities. This looks like a politically driven move (Taiwan would be expecting US military protection in exchange for the investment). For the US, it is one more step to build a self-sufficient AI supply chain (as they currently rely on Taiwan’s TSMC for actually building the chips). Also concerns grow about how this could reinforce TSMC’s position as a de-facto monopoly (e.g.: in detriment of Intel)
Like the US, China wants to be self-sufficient in chips, with Huawei as the local champion. First, as we’ve often discussed here, the commercial restrictions set by the US on Nvidia’s chips have not worked so well, and this week we have some news about how Chinese AI firms are managing to buy advanced Nvidia’s products, in spite of the bans. But these bans have also stimulated local innovation, and in particular they have accelerated the rise of Huawei as a local champion in semiconductors
Broadcom’s results this week suggest that Nvidia may have a strong competitor in AI chips. Broadcom’s presented their 1Q results this week, with better than expected revenue and a bullish outlook specifically on AI. The stock market reacted (relatively) favorably, a positive sign for a company that is worth close to $1trn. Broadcom dominates the space of ASICs (customized chips tailored for specific tasks), which are increasingly demanded by large AI companies (e.g.: Google, Meta and China’s ByteDance) to increase efficiency beyond what GPUs (more general purpose, and dominated by Nvidia) can provide
More debates about limits in the current approach to build “Artificial Super Intelligence” (ASI):
Most AI researchers don’t believe that “human-level reasoning” can be achieved with the current approach. The results of a survey published by Nature this week reveal that more than three quarters of respondents are skeptical that just building bigger deep learning models can create systems able to match or surpass human intelligence.
At the core of this problem there is a question mark on the potential of neural networks. Even if they are supposed to mimic the human brain, they are different in at least three substantial ways:
(1) artificial neurons are all functionally equivalent, while real ones are highly specialized;
(2) real networks have more complex communication patterns than artificial ones, e.g.: they use pulses of varying timing and intensity (while artificial systems forward single values across layers);
(3) real neural networks are less hierarchical, and much more interconnected than artificial ones.
Physics may play a role, at the end of the day. Part of the answer to move forward might be in a deeper integration of digital and physical systems, similar to what happens in real life. An example of this could be the work of a Princeton team that has built an ultracompact camera with optical elements that function as a neural network and can classify images at the speed of light, in a much more efficient way than existing AI-based systems (1% of the computing needed by conventional techniques)
There is a race to add AI features to consumer apps, and Apple might be losing:
This week Apple announced a delay in the launch of the new Siri app that would incorporate GenerativeAI. Meanwhile, competitors like Meta are speeding up in the development of a voice-first AI chatbot (in direct competition to Siri). This is a threat for Apple, because voice interfaces could become the “natural” way for consumer to interact with smartphones. So some analysts are suggesting a more radical approach for Apple, including ditching Siri and replace it by OpenAI’s ChatGPT
Mining is becoming trendy again, with some minerals turning into strategic assets:
This week China announced the discovery of vast deposits of thorium, a radioactive material that can be used to produce nuclear energy, potentially safer and cleaner than uranium. The deposits would be (according to the Chinese) enough to meet the nation's energy demands for 60,000 years. The finding would turn China into the global leader in thorium reserves
Rare earths are becoming a geo-strategic priority. Now at the core of the US-Ukraine discussions, this group of 17 elements are critical to modern technology and various industrial applications, including defense and aerospace (e.g.: missile guidance systems), energy storage, electronics and even medical imaging systems
Solving the problem of nuclear waste may be critical to re-accelerate nuclear energy production. US authorities are failing to permanently dispose of this waste, i.e. at sites away from nuclear reactors, mostly for political reasons. The problem is that all this could become a bottleneck for new developments. An alternative would be recycling (most of) the waste, like France (a global leader in nuclear energy) does.
Telecommunications emerge as an attractive use case for the space industry.
Starlink will play a role in US rural broadband. The Trump administration is working to facilitate Elon Musk’s Starlink platform to get access to government funding for rural broadband programs. The plan is to make the subsidies “technology neutral” instead of prioritizing fiber deployments, as they were doing under Biden
In Europe, Eutelsat may have a second life. The company’s stock had a positive reaction this week, after years of decline. The +50% jump in stock price last Monday was triggered by the company’s comments about positioning OneWeb (a subsidiary) as the European alternative to Starlink, and therefore as a tool for Europe’s self-sufficiency in technology
The promise for a healthcare revolution is emerging at the intersection between biotechnology and AI. A project launched this year by the UK Biobank genetic database and 14 drugs companies is training AI models to analyze proteins, including how they interact with other molecules (i.e.: the mechanisms behind many human diseases). The purpose is to use these models to build tools to radically accelerate the design of tailored treatments against illnesses like cancer, autoimmune diseases or dementia
Frenemies (companies competing and cooperating at the same time) appearing in the car industry:
BYD wants to cooperate with Tesla in Electric Vehicles. The Chinese company has declared its disposition to collaborate and even share key technologies in EVs and autonomous driving, in a push to fight against a “common enemy” (the internal combustion engine car). This comes in the context of (1) the recent European moves to protect local car vendors (still mostly relying on fossil fuel vehicles) and (2) the expected elimination of EV subsidies by the Trump Administration, in the US
Uber & Waymo partnering for robotaxis in Texas. The two companies are seen as big potential rivals to capture the robotaxi opportunity, with Uber dominating the user interface for taxi fleets, and Waymo dominating the autonomous car technology. Now they’re working together to make the industry grow in Austin (Texas): The Uber app will have a “Waymo on Uber” option that will make it possible for passenger to order a Waymo robotaxi
1 - Population & natural resources
Biotech
AI meets biotechnology, with a high potential impact on healthcare. A project launched this year by the UK Biobank genetic database and 14 drugs companies is training AI models to analyze proteins, including how they interact with other molecules (i.e.: the mechanisms behind many human diseases). Protein project uses AI to boost disease treatment
A startup wants to resurrect prehistoric animal species. Colossal Biosciences has raised more than $400m for this, at a valuation of $10.2bn. As a preamble to resurrecting the mammoth, they’ve produced a genetically modified mouse which has longer, thicker hair and an altered metabolism that makes it stand colder temperatures. I don’t have visibility of the business plan, but it must be fascinating ;-) Startup Trying to Un-Extinct Prehistoric Mammoth Creates ‘Woolly Mouse’
Space
Telecoms becoming a high priority use case for the space industry:
The US Dept of Commerce wants to subsidize Starlink to deliver rural broadband. They’ve modified the conditions for subsidies, previously linked only to fiber deployments. Exclusive | Commerce to Overhaul ‘Internet for All’ Plan, Expanding Starlink Funding Prospects
OneWeb looks like the “natural” candidate to be the “European Starlink”. The comments about this by the CEO of Eutelsat (which owns OneWeb) this week made the company’s share price rise +50% in one day (the starting point was not so high). Eutelsat soars as investors bet on OneWeb satellites as European option to Starlink
Consistently with this, Italy is now re-thinking its previous commitment to use Starlink (for a direct-to-cell military network). This was a $1.5bn deal, and the Italian government seems to be not so happy about the US pullback from European security commitments. Of course, Eutelsat is among the candidates to replace Starlink. Italy Is Getting Cold Feet Over Deal to Use Musk’s Starlink
Space is also being explored as an option to deploy data centers. Several startups are planning to build data centers off Earth. One of them (Lonestar Data Holdings) has just launched a computing device with 8TB of storage, which is now on the Moon, and will be in working conditions just for two weeks. Others are thinking about putting these devices in orbit. There are advantages in security, energy consumption and savings in land (obviously), but on the other side, the devices will need to withstand harsh conditions in space. Should we be moving data centers to space?
The last rocket test by SpaceX has failed. Starship, the spacecraft that SpaceX is building for missions to Mars, failed during its latest test flight on Thursday when its upper stage exploded in space. The debris from the explosion disrupted air traffic at several American airports. Breakup of SpaceX’s Starship Rocket Disrupts Florida Airports
Materials
China might have just discovered a giant deposit of thorium. This is relevant, because thorium is a radioactive material that can be used to build safer and cleaner nuclear reactors (vs. the current uranium ones). With this, China would become the country with the largest thorium reserve in the world. Map reveals where world's thorium reserves are located by country
Rare earths are becoming a geo-strategic priority. These are a group of 17 elements, which are critical to various industrial applications, including defense and aerospace (e.g.: missile guidance systems), energy storage, electronics and even medical imaging systems. Why rare earths matter to Donald Trump and the west
2 - Efficiency & Productivity
Energy
Nuclear
Solving the problem of nuclear waste may be critical to re-accelerate nuclear energy production. US authorities are failing to permanently dispose of this waste, i.e. at sites away from nuclear reactors, mostly for political reasons. The problem is that all this could become a bottleneck for new developments. An alternative would be recycling (most of) the waste, like France (a global leader in nuclear energy) does. A Nuclear-Power Revival Brings Back an Old Problem: What to Do With the Waste
Extending the life of existing reactors is a short-term opportunity, but it also has challenges. There is a global quest to squeeze more years of electricity out of existing nuclear power plants to meet rising demand for low carbon power in an efficient way. Most existing plants were built in the 80s, so they’re now reaching the end of their life times. So this plan has some risks, too (including the weakness of old building structures). Is giving old reactors new life the future of nuclear energy?
Renewables
The market for solar panels for home installations is in trouble. Sunnova, a specialist in this field, is suffering with a serious demand problem in the current context of high interest rates and lower state incentives for consumers to buy home solar equipment. Sunnova Plunges 50% After Issuing Going Concern Warning
A startup is using a new technology that vaporizes rocks to get access to geothermal energy. Quaise Energy uses a beam of electromagnetic radiation to heat the stones until they vaporize. They’re planning to drill holes of up to around 12 miles under the surface, where temperatures can reach 1,000 degrees Fahrenheit. Can a Geothermal Startup Vaporize Rock to Drill the Deepest Holes Ever?
New Transport Technologies
Electric Vehicles
BYD wants to cooperate with Tesla in Electric Vehicles. The Chinese company has declared its disposition to collaborate and even share key technologies in EVs and autonomous driving, in a push to fight against a “common enemy” (the internal combustion engine car). BYD pledges to work with rival Tesla to combat petrol cars
This happens in the context of an European effort to protect local vendors against Chinese EV suppliers. New EU tariffs are making Chinese companies like GAC review their expansion plans. Europe’s clampdown on Chinese EVs forces U-turn at state-owned GAC
Complex, costly charging infrastructure is limiting the adoption of EVs. Limited access to charging stations, especially to fast public charging, is still one of the top reasons mentioned by consumers when asked about what is limiting electric car adoption. .We’re Charging Our Cars Wrong
Autonomous Cars
Uber and Waymo are “frenemies” in Austin. A new option in the Uber app will make it possible for passengers to select a Waymo robotaxi when requesting a trip. So these two rivals will now cooperate, looking to validate the service / expand the market. Uber users in Austin are getting matched with Waymo robotaxis
Wayve, a UK-based autonomous car startup, wants to expand globally. This would be one more example of a European innovator which is able to compete with American and Chinese tech giants. UK AI start-up Wayve accelerates global expansion plans
Artificial Intelligence
AI: Apps
Agents
Meta will offer an AI customer support agents to SME customers using Instagram or Facebook as distribution channels. They’re launching a pilot program during which the service will be offered for free. Meta wants to give small businesses an AI boost with a customer support agent for Instagram and Facebook
Google deploys “agentic” features in Gemini. Users will be able to share their screens with the chatbot and ask questions about the content in them. Google's Gemini now lets you ask questions using videos and what's on your screen
B2C
Apple seems to be facing an “AI crisis”. The recent announcement of Alexa+ by Amazon (discussed here last week) is making it more urgent to update Siri… Apple’s Artificial Intelligence Efforts Reach a Make-or-Break Point
… However, the company is actually delaying the update until “the coming year”. The features were initially planned for the iOS 18.4 software update this April, but Apple engineers have been unable to fix bugs in the project. Apple Confirms Delay of AI-Infused, Personalized Siri Assistant
So some people are even suggesting them to swap out Siri with ChatGPT. The proposal would be to do this temporarily, maybe in the same way as Google Maps was the only option for maps in the first iPhone models. Apple Should Swap Out Siri with ChatGPT
Meanwhile, Apple’s rival Meta is accelerating “voice-powered AI”. The initial plan seems to be to link these features to the coming Meta AI standalone app, with voice capabilities as a component of a premium subscription, but we’ll see... This also has obvious applications for Meta’s Ray-Ban glasses. Meta accelerates voice-powered AI push
Google keeps deploying AI features in its core Search product. They’re even thinking in deploying a search-centric chatbot right to the core Google experience. Google is adding more AI Overviews and a new ‘AI Mode’ to Search
B2B
GenAI has started to impact coding jobs. As recognized by many leaders, including M Zuckerberg, AI-generated coding is a reality. This is already leading to leaner development teams and higher bars for hiring new roles. How AI Tools Are Reshaping the Coding Workforce
At many Y-Combinator startups, there are almost no coding tasks left for humans. For 25% of the Winter 2025 batch, 95% of lines of code are AI-generated. Garry Tan (@garrytan) on X
OpenAI has big plans for the education sector. This is an industry where the opportunity seems clear, and where there are significant productivity gains to be captured. OpenAI just launched a consortium that includes $50m in funding for AI projects for leading institutions. Introducing NextGenAI: A consortium to advance research and education with AI
At MWC25, AI applications for telecom operations were one of the hypes of the event. As an example of that, ServiceNow presented its new AI agents to automate operational tasks for telecoms. ServiceNow's newest AI agents bring intelligent automation to telecommunications firms
AI: Foundational Models
More debates about limits in the current approach to build “Artificial Super Intelligence” (ASI):
Most AI researchers don’t believe that “human-level reasoning” can be achieved with the current approach. This is shown by the results of a survey published by Nature this week. How AI can achieve human-level intelligence: researchers call for change in tack A Bear Case: My Predictions Regarding AI Progress
At the core of this problem there is a question mark on the potential of neural networks. Even if they are supposed to mimic the human brain, they are different in at least three substantial ways. AI versus the brain and the race for general intelligence
Physics may play a role, at the end of the day. Part of the answer to move forward might be in a deeper integration of digital and physical systems, as shown by the work of a Princeton team that has built an ultracompact camera with optical elements that function as a neural network and can classify images at the speed of light. A new way of seeing, a new way of computing | CS
Progress continues on new approaches to improve the models (albeit within the orthodox “deep learning framework”):
A Chinese team presented a paper on “Diffusion of Thought” (DoT). DoT leverages diffusion models to perform Chain-of-Thought reasoning. Rohan Paul (@rohanpaul_ai) on X
A Stanford team has analyzed under what conditions a model can gain more function through reinforcement learning. They have identified four key features that can help with this. elvis (@omarsar0) on X
Meanwhile, in spite of all the skepticism, “conventional” AI startups keep raising (lots of) money:
Safe Superintelligence, founded by ex-OpenAI scientist Ilya Sutskever, is already worth $30bn, before launching any kind of commercial offer, This Scientist Left OpenAI Last Year. His Startup Is Already Worth $30 Billion.
Reflection AI, founded by ex-DeepMind researchers, has raised $150m with the promise to build autonomous coding agents. Ex-DeepMind Researchers’ New Startup Aims for Superintelligence
Also Anthropic, a leader of the “established AI technology” has a record valuation. They just raised $3.5bn at a $61.5bn valuation. Anthropic raises $3.5 billion, reaching $61.5 billion valuation as AI investment frenzy continues
At the same time, AI has apparently stopped to catalyze Microsoft’s valuation growth. Investors are concerned about the monetization of the company’s massive AI investments. The rumored cooling down of the partnership with OpenAI also does not help. Microsoft’s Fading AI Mojo Keeps Shares in Lengthy Purgatory
“Distillation” is increasingly seen as a big promise to produce cheaper models which match the more expensive ones in capabilities. This is the methodology that DeepSeek reportedly used (starting with a larger model and refining it into a smaller, cheaper one, saving a lot of training money in the process). AI companies race to use ‘distillation’ to produce cheaper models
But there are increasing concerns about the tests used to benchmark these models. Maybe it is not so clear what “equally capable” means, anymore… Chatbots Are Cheating on Their Benchmark Tests
AI: Security & Safety
The Turing Award winners are worried about AI safety. This is sort of a “Nobel Prize” for computing, and it has been won this year by two pioneers of reinforcement learning. The first thing they seem to have done is to express their concerns about the risks of open sourcing very capable models, which they believe could turn into a safety issue. Turing Award winners warn over unsafe deployment of AI models
Meanwhile, a rather extreme article was published by Eric Schmidt and Dan Hendricks at Time Magazine, comparing AI with nuclear weapons, and suggesting the US to “treat AI chips more like uranium”. The Nuclear-Level Risk of Superintelligent AI
The article has triggered almost equally extreme reactions. An example is D Jeffries, who claims that “this is one of the worst and most dangerous proposals I have ever seen and that is saying something with the doomers” Daniel Jeffries (@Dan_Jeffries1) on X
A post this week discusses the “Alignment Problem”, i.e. the problem of aligning AI’s objectives with human values, a conventional answer by politicians and some pseudo-technical leaders to the AI safety problem. As the post shows, it is very difficult even to describe the problem in the right way… What Is The Alignment Problem?
AI: Infrastructure
In spite of all the debates about AI infrastructure needs, CoreWeave moves ahead with its IPO. The company sells a cloud infrastructure services specialized in AI (i.e. based on GPU chips). They are expecting a valuation of around $35bn, based on the same multiples applied to hyper scale cloud vendors. CoreWeave IPO puts a new twist on big tech’s ‘power law’
A problem for CoreWeave is their dependence on Microsoft, which represents 62% of the company’s revenues. That’s why recent reports that Microsoft is walking away from some of its previous commercial commitments to buy CoreWeave services don’t look too good for the IPO. Microsoft walks away from some CoreWeave commitments ahead of $35bn IPO
AI: Chips
This week TSMC announced a $100bn investment in the US, to build local chip manufacturing plans. Taiwan’s TSMC is the absolute global leader in chip production. And the company’s CEO announced this week a massive investment in the US, to build three new chip plants, two chip-packaging plants and a research and development center. Trump, Chip Maker TSMC Announce $100 Billion Investment in U.S.
Everyone sees this as a politically-driven decision. TSMC’s CEO has also said that the decision is driven by customer demand, rather than political pressures. But Trump, who hosted the press conference at the White House, presented this as a win for his current tariffs strategy. TSMC Says $100 Billion U.S. Expansion Driven by Demand, Not Political Pressure
Some analysts go so far as to link this with the Ukraine situation. According to this view, Taiwan would now be worried about a potential move by Trump to leave the country “to the mercy of China”. And this could be a way to “pay” for protection… ‘Chips on the table’: Taiwan pushes for closer US ties as China threat looms
From the US perspective, this could help build a “self-sufficient” AI supply chain. Apparently, the new factories would be mostly focused on GPUs and other chips that are critical for AI. The long-term aspiration seems to be that most of the demand of these components from US Tech Giants could be addressed with local production. TSMC plays its hand in Donald Trump’s tariff war
The announcement has clear negative implications for Intel, and that might be a risk. The company’s chip manufacturing (“foundry”) unit was a potential acquisition target for TSMC (a deal now discarded). Some voices are claiming that Trump should now “do something” about this Intel unit, because otherwise TSMC would become a de facto monopoly, which has several negative implications. Trump needs to find a solution for Intel
Broadcom’s results this week suggest that Nvidia may have a strong competitor in AI chips. .
Broadcom’s presented their 1Q results this week, and gave a bullish guidance driven by AI. The stock market reacted (relatively) favorably, a positive sign for a company that is worth close to $1trn. Broadcom shares surge on strong revenue and AI outlook
Broadcom dominates the space of AI ASICs (customized chips tailored for specific AI workloads), increasingly demanded by leading AI companies to increase efficiency beyond what GPUs (more general purpose, and dominated by Nvidia) can provide. Broadcom thumbs nose at mighty Nvidia
An example of the relevance of these chips is shown in this Meta PR, focused on how they’re using them to power the company’s Augmented Reality prototypes. Specialized Silicon: How Meta’s Custom Chips Are Revolutionizing Augmented Reality
Quantum Computing
Independent researchers have found “issues” with Microsoft’s published results about a new “Topological Qubits” approach to a scalable Quantum Computing chip. Sabine Hossenfelder reviews them in this video. Microsoft’s Topological Qubits Probably Don’t Exist, Researchers Warn
The Caltech team is also a bit skeptical. Caltech’s Jason Alicea also published a post on this topic this week. Among his conclusions, he claims that “convincing evidence of topological protection may still be far off.” What does it mean to create a topological qubit?
Intelligence Augmentation
Brain-Computer Interfaces
A Chinese-European research team has successfully used AI to decode brain signals into language. This could open a path to explore future “brain reading” applications. Generative language reconstruction from brain recordings - Communications Biology
3 - Economic / Business trends
Tech & Geopolitics
China wants to be self-sufficient in chips, and Huawei is emerging as the local champion:
The commercial restrictions set by the US on Nvidia’s chips have not worked so well, and this week we have some news about how Chinese AI firms are managing to buy advanced Nvidia’s products, in spite of the bans. Chinese Buyers Are Ordering Nvidia’s Newest AI Chips, Defying U.S. Curbs
In parallel, the bans have also stimulated local innovation, and in particular they have accelerated the rise of Huawei as a local champion in semiconductors. The US has spurred the Chinese chip industry
All these complements national efforts to become a global leader also in AI models. The end game pursued is to build a completely autonomous AI stack. Beijing Ramps Up Efforts For Tech Independence