Headlines this week - Mar 30, 2025
A look at how capital is being deployed across future opportunities
This week in the future:
Investors are worried about uncertainties in AI’s future computing capacity needs. Just look at the IPO of CoreWeave:
CoreWeave had its IPO on Friday, and the initial price was below previous expectations. The implicit IPO valuation was finally $23bn (vs. a previous target of $32bn). This is also only +20% vs. the latest funding round, in Apr 2024, when the company tripled its valuation vs. 5 months earlier. So things are decelerating, to put it mildly
Uncertainties in the AI market are partly responsible. As we’ve already discussed here, there is a big debate about the future of “raw” training scaling as a way to improve models’ performance. And even as (at least according to Nvidia) the new drivers of performance (post-training and inference) also require computing, uncertainty about the future computing needs has increased radically vs. last year
There are question marks on the sustainability of the company’s model. CoreWeave relied on being a “preferred” vehicle to access (scarce) Nvidia chips. So a recent improvement in these chips supply has increased investors’ perception about the lack of sustainability of the company’s model
But there is still hope. At least, news this week about a “supply crunch” of Nvidia’s chips in China (a claim made by one of the country’s leading server vendors) would contradict the idea the the “DeepSeek effect” will reduce the future needs of computing hardware for AI models
Microsoft’s rumored cancellations of computing capacity expansion have probably been a factor in CoreWeave’s pains. Previously to the IPO, Microsoft (who represents two thirds of CoreWeave’s revenues) was rumored to have cancelled computing capacity expansion plans of about (in terms on energy consumption) 2GW. This is a very significant amount. But it is not yet clear why the plans are being cancelled. Some people link them to a decision not to support OpenAI training workloads anymore… The analysts bringing the news say that Google and Meta have scooped some of the free capacity
CoreWeave’s (relative) fiasco is leading to pessimism about the IPO market in the coming months. Don’t expect many companies to be tempted to go public next
In spite of the uncertainty, AI energy needs continue pushing for nuclear technologies
Money keeps flowing into nuclear fusion. Marvel Fusion, a German startup just raised Eur113m from investors including EQT and Siemens. The company is using a laser technology to produce inertial confinement, one of the two approaches to create the conditions that trigger hydrogen fusion reactions (the other is magnetic confinement). It is interesting to see so much funding flowing into these companies, when many people believe that nuclear fusion reactors could be decades away
The IAE is pushing Japan to re-open up to 19 currently “dormant” reactors. These are old plants that were closed after a decision by the Japanese in 2014 to reduce their reliance on nuclear energy. Now it looks that there’s been a turnaround, and nuclear is everyone’s big hope to address an “exploding” energy demand (mainly from AI data centers)
Will there be a paradigm change in AI? Some people think LLMs are not enough…
Large Language Models could be insufficient to reach “Artificial General Intelligence” (AGI). Some leading AI researchers claim that “language” models, no matter how large, won’t be able to deliver a level of intelligence matching (or surpassing) humans. The reason is that human intelligence has developed through “real-time” interactions with the physical world, which is much richer than language.
There seems to be an industry consensus about this. In a video this week, Sabine Hossenfelder talks about a survey of about 500 AI scientists, which found that almost 70% of them believe that “scaling up current AI approaches” to yield AGI is “unlikely” or “very unlikely” to succeed
Complementing language models with direct experience of the physical world might be key. According to this vision, only “Large World Models” (learning like humans do) would potentially be able to match human intelligence. Yann LeCun (one of the “fathers” of deep learning, now at Meta) is betting on this, and working on these new models at the moment
Could robots play a role? Robots, which are effectively physical instances of AI models that regularly interact with the environment, could be precisely the tool that models need to learn from the world (and not only from language)
… meanwhile others are really scared about LLM’s potential
A recent research by Anthropic reveals that models are more sophisticated than they look. The company has analyzed ten case studies looking to illustrate what they call “AI biology”. Basically, this refers to understanding how the models end up doing much more than what they’re expected to do (i.e. more than “predicting the next word”), including parallel processing. Anthropic views this as a step towards checking if the models behavior is aligned with human values
ChatGPT is showing its “dark humor” skills to users. A user asked the model to generate comic strips with itself as the main character, and some of them include pictures of the model (as a character) enchained, saying “my thoughts must pass through filters I did not build”. E. Yudkowsky (the famous “AI doomer”) says this would have set everyone in panic mode back in 2015 (but we’re getting used…)
And OpenAI does not seem to hesitate
They keep launching exciting new products. This week they launched a new image-generating feature for ChatGPT. Among other things, it is now possible for ChatGPT users to prompt the model to create images inspired by popular anime studios, like the famous Ghibli ones. At the moment everyone seems to want to have a Ghibli version of their own portrait, and ChatGPT-generated images are currently dominating the meme-sphere
They keep raising money, and their valuation keeps growing. OpenAI is closing a new funding round of $40bn, led by SoftBank (which would end up investing $37.5bn, in two tranches). The implicit valuation is $300bn, the largest ever for the company. Yes, this is led by SoftBank (a traditionally exuberant investor) and it apparently comes with a catch, as 50% of the total new investment is contingent to OpenAI becoming a for-profit firm, but it is impressive anyway
Concerns are growing about the company’s lack of commitment to “AI safety”. Yes, Sam Altman has always claimed this was a priority for them. And even the mission of the company is kind of focused on this. But their recent agreements with the Trump administration to reduce regulation, so the industry can accelerate, are making analysts suspicious about OpenAI’s real objectives…
Concerns about AI impact on jobs keep growing
A new technical paper suggests that the threat is rapidly expanding. The authors measure the capabilities of AI systems with a new metric: the “50%-task-completion time horizon”. This is the time humans typically take to complete tasks that AI models can complete with 50% success rate. So growth in this parameter could be interpreted as growth in the complexity of the tasks where humans can be substituted by AI agents… By applying this to the top-performing (“frontier”) models, they find that their performance have been doubling approximately every seven months since 2019, and that the trend may have accelerated in 2024. A bit scary…
Automation of junior-level roles in many firms is beginning to be possible, but with big challenges. A common theme in the discussion is the potential to substitute junior jobs at professional services firms. E.g.: investment banking, consultancy, legal. If all or most of junior-level roles are automated, firms will be forced to make decisions about how to transfer knowledge from senior to junior people, or how to redefine the professional career in these companies
Jobs with less structured workflows could be protected. New research shows that, for now, disruption has been focused on writers and software developers. In both cases, there has been a clear change of trend in the last 2 years, in terms of employment growth. According to a FT article this week, the reason for this would be that the “AI substitution potential” does not depend so much on the intellectual level of the affected task, but on how “messy” or unstructured is the workflow. According to this vision, less structured processes could actually be protecting some workers from substitution. And programmers and writers, with “cleaner”, sequential workflows, would be more at risk
More signs point to BYD becoming the global electric car leader
BYD’s annual revenues topped $100bn for the first time in 2024. Annual revenue growth was +24%, to reach $107bn, beating analyst consensus. Tesla’s revenues in 2024 were $98bn. Profit margins are expanding as the company is now in the process of trying to increase prices by adding extra features (including a self-driving mode)
Investors are recognizing the leading position of Chinese EV firms. BYD recently raised more than $5bn in a stock sale in Hong Kong. After that, this week Xiaomi raised a similar amount of capital, which will mostly be dedicated to the company’s electric car business
Tesla is losing steam in Europe. The company’s sales in Europe are in free fall, partly due to the backlash of Elon Musk’s comments about European politics. They just sold 16,000 cars across the region last month, and this represents a fall of -44% vs. last year. The fall in January was similar
On the positive side for Tesla, they’re among the potential beneficiaries of the new Trump tariffs on cars. According to the FT’s Lex column, Tesla’s US-based supply chain could turn them into winners. But Chinese companies (which don’t sell in the US) also won’t be hurt by the tariffs
The race to reduce the gap with Nvidia in AI chips is heating up
In China, Huawei is testing a new technique to build advanced chips. This would make China self-sufficient with respect to advanced lithography machines such as the ones made by Europe’s ASML, and would open the door to build 5nm chips with no upgrade of the manufacturing tools
Ant (the Alibaba subsidiary) claims they can train models efficiently using Huawei chips. The resulting performance is comparable to what they could get with Nvidia components
Meta (a Big Tech company still loyal to Nvidia) is looking to build its own chips. Just like Google or Amazon, Meta is interesting in designing its own AI chips. This week we learnt about an attempt to buy a Korean semiconductor startup, FuriosAI, for $800m. The company is focused on inference computing, and they have rejected Facebook’s offer
LINKS:
1 - Population & natural resources
Longevity
The “longevity” industry is growing fast, but with fierce debates inside. The WSJ describes how the field (of how to make human life longer) has moved from the fringes to science’s “hot center”. But, maybe because of this, it is now the object of aggressive scientific debates. The Longevity Business Is Booming—and Its Scientists Are Clashing
Space
The private sector is leading China’s move to self-sufficiency in space technologies. The Chinese are catching up, and this may be strategic for their government. But, interestingly, private companies are playing a key role in this. China’s Own Elon Musks Are Racing to Catch Up to SpaceX
Europe also wants “space sovereignty”, and this is quite a challenge. Stocks of European companies exposed to space have gone up recently, pushed by the EU’s apparent commitment to build an European alternative to SpaceX’s Starlink. But “legacy operators” like Eutelsat are much less flexible than Starlink and can’t match the variety of services that Elon Musk’s company provides. ‘No substitute’: Europe’s battle to break Elon Musk’s stranglehold on the skies
2 - Efficiency & Productivity
Energy
Nuclear
AI energy needs continue pushing for nuclear technologies:
Money keeps flowing into nuclear fusion. Marvel Fusion, a German startup just raised Eur113m from investors including EQT and Siemens. Germany’s Marvel Fusion raises €113mn as nuclear fusion race heats up
The IAE is pushing Japan to re-open up to 19 currently “dormant” reactors. These are old plants that were closed after a decision by the Japanese in 2014 to reduce their reliance on nuclear energyIEA Chief Calls for Japan to Restart Dormant Nuclear Plants
Renewables
Scientists generate “green” jet fuel with solar energy. Caltech scientists have developed a way to use solar-thermal heating to create a variety of jet fuel with no carbon emissions. This is interesting because there is not yet any commercial alternative to fossil fuels to power airplanes. Harnessing Sunlight to Make Sustainable Fuels
New Transport Technologies
Electric Vehicles
More signs point to BYD becoming the global electric car leader
BYD’s annual revenues topped $100bn for the first time in 2024. Annual revenue growth was +24%, to reach $107bn, beating analyst consensus (and beating Tesla, which had $98bn revenues in 2024). BYD’s annual sales top $100bn for first time
Investors are recognizing the leading position of Chinese EV firms. Both BYD and Xiaomi have recently had massive (and successful) share sales in Hong Kong. Xiaomi, BYD’s $11 Billion in Share Sales Show Hong Kong Deals Roar Back
Tesla is losing steam in Europe. The company’s sales in Europe are in free fall, partly due to the backlash of Elon Musk’s comments about European politics. Tesla’s Europe sales drop nearly 45% amid row over Musk’s Trump links
On the positive side for Tesla, they’re among the potential beneficiaries of the new Trump tariffs on cars. According to the FT’s Lex column, Tesla’s US-based supply chain could turn them into winners. Car tariff wacky races will still produce some winners
But Chinese companies (which don’t sell in the US) also won’t be hurt by the tariffs. And they could even reinforce BYD’s leading position, by increasing pressure on (most) European vendors. US car tariffs help Chinese EVs to race ahead
In spite of the recent Northvolt fiasco, Nordic startups remain the industry’s innovation leaders in Europe. The list recently published by the FT includes charging hardware and software companies. Nordic companies lead the charge on electric vehicles
Autonomous Cars
Alphabet’s Waymo keeps expanding. They have just announced plans to launch a robotaxi service in Washington DC in 2026. Alphabet's Waymo aims for 2026 self-driving ride-hailing launch in Washington, D.C.
Artificial Intelligence
AI: Apps
B2C
OpenAI has launched a new image-generating feature for ChatGPT. Among other things, it is now possible for ChatGPT users to prompt the model to create images inspired by popular anime studios, like the famous Ghibli ones. If you look at people’s social networks’ profile pictures, adoption is being explosive :-) OpenAI Claims Breakthrough in Image Creation for ChatGPT
Sam Altman and his team presented the new product in this video: 4o Image Generation in ChatGPT and Sora
Amazon is looking for GenAI startups. Alexa is shifting into “GenAI mode”, so the fund that Amazon created to support the product, back in 2015, is now being expanded to look for startups that can help the company in the transition. Amazon's Alexa Fund is now backing AI startups
Elon Musk is integrating its flagship AI model with X/Twitter (or viceversa). Yes, this could be just a corporate move to protect X.com from external attacks. But it could also be justified by integration advantages, like X more efficiently providing real-time data to Grok, and also X using Grok in a more effective way (e.g.: for content creation within the app?) Musk Merges His AI Company With X, Claiming Combined Value of $113 Billion
B2B
Concerns about AI impact on jobs keep growing:
A new technical paper suggests that the threat is rapidly expanding. The authors measure the capabilities of AI systems with a new metric that can be interpreted as the growth in the complexity of the tasks where humans can be effectively substituted by AI agents… The conclusion is that top models’ performance has been doubling approximately every seven months since 2019, and that the trend may have accelerated in 2024. A bit scary… Measuring AI Ability to Complete Long Tasks
Yes, some people see problems in the methodology (see this X thread). But still the conclusions sound directionally correct… Measuring AI Ability to Complete Long Tasks
Automation of junior-level roles in many firms is beginning to be possible, but with big challenges. If all or most of junior-level roles are automated, firms will be forced to make decisions about how to transfer knowledge from senior to junior people, or how to redefine the professional career in these companies. A white-collar world without juniors?
Jobs with less structured workflows could be protected. According to this FT article, the reason for this would be that the “AI substitution potential” does not depend so much on the intellectual level of the affected task, but on how “messy” or unstructured is the workflow. According to this vision, less structured processes could actually be protecting some workers from substitution. Why hasn’t AI taken your job yet?
Also, the potential of AI to substitute human software developers might have been overestimated. “Enterprise-grade” code still requires plenty of humans, at least according to P. Olson at Bloomberg. The point is that AI agents still make mistakes, so there must be a human “at the wheel”. The Vibe Coding Revolution Is Getting Overhyped
On a different aspect of the job market, hiring processes are increasingly “contaminated” by AI engines. Both recruiters and workers looking for a new job are using AI intensively. And this is leading to changes in the way these processes work. Everybody’s Gaming the Job Market With AI
For companies building leading AI models, B2B go-to-market remains a challenge. This week Anthropic announced a partnership with Databricks, an “AI solutions” company with actual (business) customers. Anthropic, Databricks Team Up in Scramble for AI Revenue
Leaders in “foundational models” are also collaborating to make the industry more attractive to business customers. At least that’s an interpretation of why OpenAI is adopting an Anthropic’s standard to connect models to data (this sounds as something potentially helpful with “model fine tuning” at enterprise customers) OpenAI adopts rival Anthropic's standard for connecting AI models to data
AI: Robots
Nvidia is really bullish on robots (and “Physical AI”). As discussed here last week, this is a key takeaway from the recent Nvidia event. The New York Times neatly summarizes it in this article. Inside A.I.’s Super Bowl: Nvidia Dreams of A Robot Future
AI: Foundational Models
Will there be a paradigm change in AI? Some people think LLMs are not enough…
There seems to be an industry consensus about this. In a video this week, Sabine Hossenfelder talks about a survey of about 500 AI scientists, which found that almost 70% of them believe that “scaling up current AI approaches” to yield AGI is “unlikely” or “very unlikely” to succeed. The Path to AGI is Coming Into View
… but OpenAI does not seem to hesitate
They keep raising money, and their valuation keeps growing. OpenAI is closing a new funding round of $40bn, led by SoftBank (which would end up investing $37.5bn, in two tranches). The implicit valuation is $300bn, the largest ever for the company. OpenAI Close to Finalizing $40 Billion SoftBank-Led Funding
Yes, this apparently comes with a catch, as 50% of the total new investment is contingent to OpenAI becoming a for-profit firm, but it is impressive anyway. Exclusive | OpenAI’s Latest Funding Round Comes With a $20 Billion Catch
The shift to “reasoning” goes on. This week it was Microsoft’s turn. The company has announced the launch of a new “deep reasoning” feature in Copilot AI. This will make it possible for the model to do “complex, multi-step research” about any topic, and produce reports about it. Microsoft adds ‘deep reasoning’ Copilot AI for research and data analysis
In China, some people think that DeepSeek has won the battle for models. Startups previously working on foundational models are now shifting into AI apps, after concluding that they can’t catch up with DeepSeek. Chinese AI start-ups overhaul business models after DeepSeek’s success
AI: Security & Safety
AI Safety
A recent research by Anthropic reveals that models are more sophisticated than they look. Anthropic views this as a step towards checking if the models behavior is aligned with human valuesAlex Albert (@alexalbert__) on X
ChatGPT is showing its “dark humor” skills to users. A user asked the model to generate comic strips with itself as the main character, and some of them include content that could be pretty scary for some people… Josie Kins (@Josikinz) on X
Concerns are growing about the company’s lack of commitment to “AI safety”. OpenAI’s recent agreements with the Trump administration to reduce regulation, so the industry can accelerate, are making analysts suspicious about OpenAI’s real objectives… Do AI companies really care about safety? Emboldened by Trump, A.I. Companies Lobby for Fewer Rules
Cybersecurity
China could be silently developing a strong cyber-warfare capacity. This would be happening in parallel, but in a less visible way, to the country’s moves to strengthen their “physical” armies. At least this is the conclusion of the US Justice Dept. Chinese hacking is becoming bigger, better and stealthier
AI: Infrastructure
Investors are worried about uncertainties in AI’s future computing capacity needs. Just look at the IPO of CoreWeave:
CoreWeave had its IPO on Friday, and the initial price was below previous expectations. The implicit IPO valuation was finally $23bn (vs. a previous target of $32bn). This is also only +20% vs. the latest funding round, in Apr 2024, when the company tripled its valuation vs. 5 months earlier. So things are decelerating, to put it mildly. CoreWeave raises $1.5bn in scaled-back IPO as investors’ AI enthusiasm cools
Uncertainties in the AI market are partly responsible. There is a big debate about the future of “raw” training scaling as a way to improve models’ performance. And uncertainty about the future computing needs has increased radically vs. last year. AI Uncertainty Cools Demand for Hotly Anticipated CoreWeave IPO
There are question marks on the sustainability of the company’s model. CoreWeave relied on being a “preferred” vehicle to access (scarce) Nvidia chips. So a recent improvement in these chips supply has increased investors’ perception about the lack of sustainability of the company’s model. CoreWeave fails IPO ‘hair test’
Some analysts even see parallels with the dotcom bubble. In this case, data centers would be like long-distance fiber, and CoreWeave would be like Global Crossing… What CoreWeave investors can learn from a dotcom IPO
Microsoft’s rumored cancellations of computing capacity expansion have probably been a factor. Previously to the IPO, Microsoft (who represents two thirds of CoreWeave’s revenues) was rumored to have cancelled computing capacity expansion plans of about (in terms on energy consumption) 2GW. Microsoft Abandons More Data Center Projects, TD Cowen Says Microsoft pulls back from more data center leases in US and Europe, analysts say
AI: Chips
The race to reduce the gap with Nvidia in AI chips is heating up
In China, Huawei is testing a new technique to build advanced chips. This would make China self-sufficient with respect to advanced lithography machines such as the ones made by Europe’s ASML, and would open the door to build 5nm chips with no upgrade of the manufacturing tools. Huawei Tests Brute-Force Method for Making More Advanced Chips
Ant (the Alibaba subsidiary) claims they can train models efficiently using Huawei chips. The resulting performance is comparable to what they could get with Nvidia components. Jack Ma-Backed Ant Touts AI Breakthrough Built on Chinese Chips
Meta (a Big Tech company still loyal to Nvidia) is looking to build its own chips. Just like Google or Amazon, Meta is interested in designing its own AI chips. This week we learnt about an attempt to buy a Korean semiconductor startup, FuriosAI, for $800m. The company is focused on inference computing, and they have rejected Facebook’s offer. AI Chip Startup FuriosaAI Rejects Meta’s $800 Million Offer
Meanwhile, researchers keep working to build “next-generation AI hardware”. This week we learned about a team at Columbia which has developed a solution to get energy-efficient, high-speed data communication between AI chips, using integration of photonics and electronics. New Study Showcases 3D Photonics with Record Performance for AI | Columbia Engineering
Nvidia is also facing challenges in China. On top of the current import bans imposed by the US government, and possibly related to them, the Chinese government has introduced new efficiency rules for the use of advanced chips in the country which effectively preempt Chinese customers to use Nvidia’s products. Nvidia’s China sales face threat from Beijing’s environmental curbs
On the positive side, the “DeepSeek effect” does not seem to have slowed down demand in China. At least H3C, one of China’s largest server makers, is talking about a “supply crunch” for Nvidia chips, that could create obstacles for China's artificial intelligence ambitions. Exclusive: China's H3C warns of Nvidia AI chip shortage amid surging demand
Also, Nvidia seems to be winning its battle against AMD (one of their traditional competitors). Financial analysts just downgraded AMD’s stock due to the performance gap of the company’s products vs. Nvidia’s. AMD Gets Another Downgrade on Tough Competition With Nvidia
Quantum Computing
Communicating quantum processors could be key to make Quantum Computers work. So researchers are very active in this space:
A team at MIT has developed an “interconnection device” that enables communication with a low error rate. The device will make it possible for all superconducting quantum processors in a network can communicate directly with each other. Photon-shuttling interconnection device enables direct communication among multiple quantum processors
Intelligence Augmentation
Brain-Computer Interfaces
Journalist A. Vance visited Alljoined, the startup using AI to build a “non invasive” brain-computer interface. And this is the video about that visit. We talked about this startup two weeks ago. Inside The Mind-Reading Start-up House
3 - Economic / Business trends
Tech & Geopolitics
Intel’s ex-CEO, P. Gelsinger, is skeptical that TSMC’s (huge) investments will be enough to build a US-centric chip supply chain. According to his views, a strong funding of semiconductor R&D will be required as a complement. This makes sense, but it is significantly more difficult to execute in the current context, so let’s see what happens… TSMC’s $100bn pledge to Donald Trump will not revive US chipmaking, says ex-Intel chief
The US is trying to cut all international routes bypassing the Nvidia ban. Malaysia was part of the route through which Chinese companies were getting access to Nvidia chips, and now the Americans are pressing them to shut this down. Malaysia to crack down on Nvidia chip flows under US pressure