Market in 2026: The Evolving Role of AI

My initial idea was to make my first post of the year a longer piece covering the current market environment, the overall outlook for 2026, the leading investment theme for the year, and a few other related topics. As I started writing, however, I realized I had much more to say about the market itself, and that section naturally evolved into a post of its own.

Now that we’ve clarified my views on the current market environment and the broader picture for the year, let’s move into more uncharted territory in this second part by exploring what I believe could be the leading investment theme for the years ahead.

AI Remains the Core Theme—With a Shift

In the summer of 2024, at a time when many were loudly arguing that we were already in an unsustainable AI bubble—one reminiscent of the dot-com era and on the verge of bursting—I felt it was the right moment to take a closer look at what actually happened between 1995 and 2001. One of the key conclusions from that exercise was that, if the dot-com bubble were indeed a useful map for how an AI bubble might unfold, we were still very far from the 2000 blow-off phase. If anything, the more appropriate analogy at that time was 1996.

Back then, it was already clear that the internet was going to be transformative, but how that transformation would ultimately monetize and gain widespread traction was still uncertain. As a result, the early winners of that period were companies like Cisco—firms that were providing the core technology on which the internet revolution was being built.

This closely mirrors what we observed from 2023 through 2024. While investors were captivated by AI’s long-term potential, there was still limited clarity on how it would concretely transform everyday life and business models. As a result, the early winners were not end-user applications, but rather the companies providing the technological backbone of the AI revolution. AI chipmakers such as Nvidia, memory producers like Micron, cooling solution providers like Vertiv, energy companies, and other infrastructure-focused players emerged as the most obvious beneficiaries. On the user side, the ultimate winners were still unclear, leaving infrastructure as the only truly visible and investable theme at that stage.

At the time, I argued that—much like the rally surrounding the internet as a commercial platform in 1997—the real euphoria around AI would only emerge once the technology entered its speculative, transformative phase. That phase would begin when AI started to meaningfully reshape a broad range of business models, lift productivity across multiple sectors, and allow entirely new leaders to emerge—rather than confining the upside to companies merely building the underlying infrastructure.

I’m summarizing this view here because I believe 2026 may be the first year in which that transition becomes really visible.

That conviction is not based on the idea that this is simply the next logical step in AI’s evolution, but rather on what is concretely brewing right now.

If your primary use of large language models has been text editing, summarizing, or writing, the past year may have been somewhat deceiving in terms of how you perceived AI progress. For these use cases, the truly explosive phase ran from 2022—when ChatGPT was impressive but deeply imperfect—to 2024, when it became genuinely solid. From that perspective, ChatGPT-4o already felt close to a peak. This is likely why many users were frustrated by the initial release of ChatGPT-5, and why, even today with a very strong ChatGPT-5.2, some people still nostalgically miss GPT-4.5 when it comes to writing tasks.

However, if you were more demanding in the diversity of AI applications—using it not just to write, but to actually do things such as complex research, report generation, PowerPoint creation, or coding—then 2025 must have felt like an extraordinary year. I genuinely don’t think I’ve ever witnessed such a combination of massive technological improvement occurring simultaneously with a collapse in the cost and friction of integration. That dual effect stems from a simple but profound dynamic: as AI models became better at coding, they also became better at coding their own integration.

As recently as March–April, when Zack and I were working on WU Advanced and TuneMap, we were using the GPT-o3 reasoning model, which at the time was the most advanced available. For us, it was eye-opening. It was the first time AI truly allowed us to code faster in a meaningful way. The generated code often didn’t compile or didn’t quite do what we wanted, but it still accelerated us to the point where we effectively operated like a team twice our size.

Then, in September, when Claude Sonnet 4.5 launched—and later Opus 4.5 in December—we crossed into something of an entirely different magnitude. These models can now create complex projects in ways I simply could not have imagined before. Some studies suggest that an average developer writes around 250–300 lines of code per day. Today, we can easily generate 2,000 lines per day. PandAI alone probably represents something like 20,000 lines of code, and we were able to build it in a surprisingly short amount of time.

We even turned this into a running joke in our repository. One morning, one of us created a folder called “Pandanoide V1” with a note saying “to be improved.” We are now at Pandanoide V4—a modern reinterpretation of the Atari game Asteroids, where you pilot a giant panda spaceship that can summon a massive manga-style Tom Lee to obliterate everything after collecting five Bitcoins, while a well-known analyst also occasionally appears to throw Elliott Wave lightning bolts to defend the ship. (Someday it will be embedded as an Easter egg into Pand-AI under a specific ticker.)

Coding has become so accessible that Andrej Karpathy, one of the co-founders of OpenAI, remarked this year that “the hottest programming language is English.” I completely agree with that statement. Even Linus Torvalds, the creator of Linux and a near-mythical figure in the world of software development, recently admitted that he is guilty of “vibe coding”—using AI as the primary tool to build software projects.

AI is still imperfect and will obviously continue to improve. But I’m emphasizing all of this because I believe we’ve now reached a point where AI can do genuinely serious work—and, crucially, can deploy itself inside organizations with remarkable ease. Beyond the progress of the models themselves, several key technologies around AI integration have also improved dramatically. For the more technical readers: retrieval-augmented generation has become both easier to implement and far more efficient, making fine-tuning models almost unnecessary in many cases.

To give a sense of scale, in 2024 a well-known company approached me about integrating a chatbot into Robotiq’s internal documentation. They framed it as a one-year, one-million-dollar project. Today, you could build a fairly functional version of that in two to three weeks.

This is why we’re starting to see AI features everywhere: on Yahoo Finance, on Amazon (ask Rufus), across the entire Google ecosystem (Drive, Gmail, and more), on Expedia, and countless other platforms. From my point of view, this is only the beginning. We are probably going to get sick of seeing AI everywhere, endlessly generating text down our throats. But the companies that manage to make AI genuinely useful—or deploy it transparently to meaningfully boost productivity rather than as a superficial gizmo—are the ones that will see their revenues and/or profits increase significantly, and, as a result, their valuations meaningfully appreciate.

This brings me to my first market prediction. Unlike the ones that will follow, this one is less centered on specific names I think will win and more on a broader market dynamic. Fueled by this trend—and accelerated by lower interest rates—I believe the market will finally broaden out after spending nearly four years on the bench. My guess is that we will revisit that zone at some point in 2026. At that stage, it will no longer be just about Big Tech building AI infrastructure, but about a much wider set of companies actively embracing and deploying AI across their businesses.

CoreWeave and Nebius could continue to appreciate, no longer merely as speculative plays, but as direct beneficiaries of real AI adoption by companies. Snowflake, which helps organizations prepare, query, and serve the data that feeds AI models and applications—often via APIs, connectors, and data pipelines—could also have a strong year if it manages to shift the trajectory of its EPS.

I will also continue to bet on UiPath, which was one of my positions last year. As a leader in intelligent document processing and AI-augmented software testing tools—and with partnerships spanning Nvidia, OpenAI, Microsoft, and Snowflake—this company, which currently carries no debt, could perform very well if execution follows.

That said, these are still primarily plays on AI adoption itself. The real winners of the adoption wave, as I’ve argued, will more likely be the companies that use AI to materially boost productivity, expand revenue, and create entirely new products that achieve rapid adoption. Those winners are difficult to identify at this stage, but this is precisely what I’ll be watching closely with Pand-AI, quarter after quarter, every earnings season.

IPOs of young, AI-driven disruptors could also prove very exciting. Lower barriers to integration mean a much larger pool of potential disruptors—but that same dynamic should also create compelling short opportunities. Legacy companies that fail to adapt risk being left behind by AI. This is not speculative. Have you seen what happened to Stack Overflow in December?

For developers, Stack Overflow was the social platform. Everyone used it. Whenever I was coding in Pine Script (TradingView), C, or Python, I always had multiple Stack Overflow tabs open in my browser. There was even a running joke a few years ago that coding mostly consisted of stitching together chunks of code found on Stack Overflow—which, honestly, wasn’t far from the truth.

Not long ago, Stack Overflow was seeing around two hundred thousand posts per month. In December, as developers migrated en masse to AI-based tools, the number of questions asked collapsed to just 3,862. That represents a staggering 98% decline.

We should expect similar stories to play out across many other industries and business models.

AI goes physical

In January 2025, Jensen Huang, CEO of Nvidia, said that after agentic AI, the next logical step for artificial intelligence would be physical AI. But what does “physical AI” actually mean?

In the fall of 2022, ChatGPT—combined with DALL·E 2—stunned the world by demonstrating that machines could generate text and images at a level that rivaled human creativity and abstraction. Almost overnight, entire categories of work suddenly felt exposed: illustrators, marketers, writers, journalists, teachers. Over the following years, AI continued to improve, pushing further into these domains. Then agentic AI unlocked an entirely new class of capabilities, allowing models not just to generate content, but to plan, execute, and complete tasks autonomously. As a result, new professions came into focus: software developers, financial analysts, paralegals, back-office operations, HR, software testing, project management, and more.

What all these roles have in common is that they are desk jobs, mediated by a computer, and heavily reliant on abstraction.

At the same time, jobs grounded in physical interaction—manipulation, dexterity, contact, force, and perception—remained largely untouched. That contrast surprised many people, including myself (well not so much).

I vividly remember a day in November 2022 that made this gap impossible to ignore. In the morning, Jennifer (WU co-founder) and I were experimenting with ChatGPT and DALL·E 2 to generate marketing text and visuals for WealthUmbrella. We were genuinely amazed by how well the system handled abstract reasoning, language, and creative synthesis. That same afternoon, I was sitting on a master’s thesis defense. The student had spent two years developing a more efficient strategy for inserting a pin into a hole with a robot—a task that any human performs effortlessly, yet one that modern robots still struggle with deeply.

That single day crystallized a striking reality: machines had learned to rival us in abstract reasoning and creativity before they could reliably perform some of the simplest physical tasks a human takes for granted.

This apparent contradiction is not new. It was already identified in the 1980s through what is known as Moravec’s paradox. Formulated by robotics researcher Hans Moravec, the paradox observes that tasks humans find easy—such as perception, movement, and intuitive manipulation—are extraordinarily difficult for computers, while tasks humans find difficult—like logic, mathematics, or chess—are comparatively easy for machines. The reason is deeply rooted in evolution: our sensory and motor skills are the result of millions of years of adaptation, whereas abstract reasoning is a relatively recent development in human cognition.

Another way to frame it is that abstract reasoning is largely conscious—we can articulate it, formalize it, and teach it—while physical intelligence is learned implicitly and operates beneath awareness. We know how to pick up an object, adjust grip force, or align a pin with a hole, but we struggle to explain how we do it. That makes physical intelligence far harder to replicate in machines.

This is precisely why physical AI matters—and why it is now emerging as the next major frontier. If AI want to continue to disrupt job market it have to bridge the gap into physical world. And in that domain we have very liittle knowledge. Indeed, one of the reason LLM beca,e rapidly clever was in part due to the maount of written data we had accumulated. Particularly a big thanks to Google which had for more than a decade digitzalized so many of the stuff that mankind had written over the century. Generative AI was build over that, but there is no equivalent in the physical domaine. So the next logical step in this new world is data collection.

Around mid-2025, several AI companies began building what can best be described as robot farms—large-scale robotic setups designed specifically to collect physical interaction data. Conceptually, it looks something like this:

(Source: Agile Robots)

I know this not from theory, but from direct exposure. I’m one of the owners of the company that sells the most popular robotic gripper in the world. I obviously can’t disclose specifics without breaching customer confidentiality or our own business intelligence, but I can say this plainly: the scale and intensity of what is happening right now is unlike anything I’ve seen before. This is not a side experiment. It is a full-blown shift.

Collecting physical data is the foundation for building physical intelligence—but data alone isn’t enough. You also need a body to act, interact, fail, learn, and repeat. And this is where things become even more striking.

An almost absurd amount of capital is now flowing into humanoid robotics.

So much so that the Chinese government has publicly warned about the risk of a humanoid-robot bubble. In that context, venture funding in this space has exploded. A few examples illustrate just how serious this has become:

1. 1X Technologies (OpenAI-backed humanoid robotics)

   The humanoid robotics company 1X Technologies, backed by the OpenAI Startup Fund, has been rapidly expanding its fleet and has publicly stated plans to produce tens of thousands of humanoid units by 2026. Crucially, their training strategy is shifting toward robots learning directly from their own video streams and physical interactions, reducing reliance on human teleoperation. This is data collection at scale, embedded directly into deployment.

2. Physical Intelligence

   The robotics AI startup Physical Intelligence has raised hundreds of millions of dollars from investors including Alphabet’s venture arm and Jeff Bezos. Their explicit goal is to train foundation models on real-world robotic manipulation data. Rather than relying primarily on simulation, they are grounding intelligence in physical experience.

3. Genesis AI

   Genesis AI, a full-stack physical-AI research lab, emerged from stealth with over $100 million in funding. Their mission is to build a universal robotics foundation model alongside a scalable data engine designed specifically to ingest and learn from real robotic interaction data.

Taken together, these initiatives point to the same conclusion: physical AI is no longer a speculative concept. The industry has moved past debating whether physical intelligence matters and is now racing to collect the data, build the bodies, and establish the foundations required to make it real.

I believe the best investment opportunities in physical AI for 2026 will still largely sit in the private capital domain. That said, I’m equally convinced that this trend will increasingly spill over into public markets.

A recent and very telling example is SoftBank Group (SFTBY). SoftBank acquired ABB Robotics for close to $6 billion—not through its futuristic Vision Fund, but directly on its balance sheet. This was an explicit signal that robotics and physical AI are now considered part of SoftBank’s core operations. They publicly labeled the transaction a Physical AI play and the market reaction was immediate: SoftBank’s stock jumped roughly 13%. At SoftBank’s market capitalization, that single-day move represented a valuation increase roughly five times larger than the price they paid for ABB Robotics itself.

SoftBank also owns roughly 20% of Boston Dynamics, which arguably operates the most impressive and dynamically capable humanoid platform in the world. Given this positioning, I would not be surprised to see SoftBank continue making aggressive moves in this space. It is a company I will be watching closely in 2026.

Beyond SoftBank, here are other public companies that could benefit meaningfully from the physical AI and humanoid trend.

Tier 1: the “picks and shovels” winners

This first tier is, in my view, the least controversial and the safest way to gain exposure to humanoids and physical AI. These companies win almost regardless of which humanoid platform ultimately succeeds.

NVIDIA (NVDA) is the most obvious example. At this point, it is the clear “brain supplier” for the entire ecosystem. Its Jetson Thor platform and Isaac GR00T models are being adopted by nearly all serious humanoid developers, including Boston Dynamics, Figure AI, Agility Robotics, 1X Technologies, Apptronik, and others. In January 2026, NVIDIA even released physical-AI models explicitly designed for next-generation robots. This is classic infrastructure dominance: wherever humanoids go, NVIDIA goes with them.

Texas Instruments (TXN) fits perfectly into the same category, but on the less glamorous side of the stack. TI supplies analog chips, precision sensors, and motor controllers that effectively form the nervous system and muscles of robots. They recently partnered with KUKA to integrate their real-time motor-control and sensing technology directly into industrial robot arms. This is a textbook pick-and-shovel play: not flashy, but deeply embedded.

Analog Devices (ADI) occupies a very similar position. Precision sensing, motion control, and signal processing are absolutely critical for physical AI, and ADI quietly benefits every time robots move out of labs and into real-world deployment.

Tier 2: component specialists that tend to be overlooked

This tier is often ignored because these companies don’t “look” like AI plays, but they are structurally indispensable.

Nabtesco (6268.JP) is a perfect example. The company controls roughly 60% of the global market for precision reducers used in industrial robots. Every articulated robot arm needs these components. There is no humanoid revolution without them.

Sony (6758.JP) is another underappreciated player. Sony dominates the global image-sensor market, and image sensors are, quite literally, the eyes of robots. Even as models improve, high-quality vision hardware remains essential.

Hesai Technology (HSAI) also deserves mention. Most humanoid platforms—Tesla being the notable exception—still rely on LiDAR alongside cameras for spatial awareness and navigation. As long as that remains true, LiDAR suppliers stand to benefit quietly but meaningfully.

Mobileye (MBLY) is also a beneficiary of this broader physical-AI trend. While best known for autonomous driving, Mobileye sits at the intersection of perception, decision-making, and real-world deployment—core elements of physical intelligence. As markets begin to seriously price in automated vehicles, Mobileye could see renewed appreciation. In my view, this is one of the cleaner public-market ways to play embodied AI outside of humanoids.

Tier 3: integrated platform players

This is where things become more interesting

Tesla (TSLA) is the obvious high-risk, high-reward bet. Tesla’s ambition is to bring Optimus to a $20,000–$30,000 price point, compared with roughly $140,000 for Boston Dynamics’ Atlas. If—and this is a very big “if”—Tesla succeeds at manufacturing scale, the implications would be transformative. Execution risk, however, remains very real.

Hyundai Motor Group (HYMLF) also deserves attention. Hyundai owns a majority stake in Boston Dynamics and has announced plans to manufacture up to 30,000 robots annually by 2028, with Atlas units deployed directly into its own factories. This is physical AI being adopted internally before being commercialized externally, which is often a strong signal.

Tier 4: the application layer

This is where physical AI stops being a technology story and becomes a business story.

Intuitive Surgical (ISRG) is the gold standard. With more than 10,000 da Vinci systems deployed globally and recurring revenue tied to procedures, Intuitive has already proven how to monetize physical AI in healthcare at scale.

Symbotic (SYM) is another strong example. Its multi-year partnership with Walmart to deploy AI-driven warehouse robotics across hundreds of locations demonstrates how quickly productivity gains can compound once physical AI is embedded into core operations.

Amazon(AMZN) is already operating more than one million robots across its fulfillment network—the largest deployed robotics fleet in the world. This is not a future bet; it is already happening.

Taken together, this entire stack—from chips and sensors to platforms and real-world applications—is why I believe physical AI is no longer a speculative concept. It is moving quietly but decisively into deployment, and the market is only just beginning to price that in.

Other trends

I also don’t see how energy, copper, and silver would not continue their upward trend. These are not speculative themes, like cannabis was in 2018. This is driven by very real and growing demand, and I see no credible reason for that demand to diminish.

The number of electric motors and electronic circuits being produced globally is increasing at a pace that far exceeds the growth in copper and silver production. As a result, these materials are increasingly becoming scarce. There is nothing on the horizon suggesting this dynamic will reverse.

AI adoption only reinforces this trend. The computing power required to train and run modern AI models is enormous, and that demand continues to rise. All of this computation requires energy—an extraordinary amount of it—and energy itself is becoming a constrained resource.

Many people are aware of how much electricity Canada produces thanks to its massive hydroelectric capacity. A significant portion of that capacity is located in the province of Quebec, where I live. One of the major advantages here has historically been access to cheap electricity. That said, even here, energy is no longer abundant. The few remaining surplus megawatts are now being reserved by the government for strategic projects.

If energy has become constrained in Quebec, it is even more constrained almost everywhere else on the planet—especially as AI is gobbling it up like the blue monster from Sesame Street devouring cookies. In that context, energy clearly exhibits all the characteristics of a long-term megatrend.

Bitcoin

Bitcoin didn’t have its best year in 2025. Technically, it ended the year about 5% lower than where it started. In between, it experienced its usual ups and downs, but even the upside moves were relatively muted by Bitcoin standards—roughly 25% from the start of the year at the highs.

That said, it would be wrong to call 2025 a bad year. Historically, bad Bitcoin years tend to be far more violent than that. So here is my view on Bitcoin for 2026.

Not a cyclical downtrend

Bitcoin has declined meaningfully since peaking in early October 2025. The primary driver of that move was a massive outflow from Bitcoin ETFs. In hindsight, this was inevitable. Those ETFs were the main catalyst of the cycle, and with every major buzz comes a return to a more normalized state. In that context, having Bitcoin trading close to $100,000 after this normalization phase is actually quite bullish from a long-term perspective. The ETF-driven excitement could have unwound far more violently.

That said, our long-term view has not changed. None of our market-top indicators—each looking at Bitcoin from very different angles—has come anywhere close to readings consistent with a euphoric top. This aligns well with what we observed throughout 2025: Bitcoin never really entered a euphoric phase.

Many market participants who still believe Bitcoin strictly follows a four-year halving cycle have labeled this correction a cyclical downtrend. As I’ve said repeatedly, we already argued in 2023 that Bitcoin would likely stop following that rigid four-year pattern and that its timeline could become very different. For that reason, we warned against relying too heavily on cyclical timing strategies and instead focusing on overbought and oversold dynamics.

The rationale was simple: as mining rewards continue to shrink and new supply becomes marginal relative to the existing circulating supply, the halving should have a diminishing impact on price swing. And that is exactly what we’ve observed. Nearly two years after the halving, we have not seen an euphoric, supply-driven rally. If there was a true catalyst behind the stronger price action, it was ETFs—not the halving.

For us, what Bitcoin has been experiencing since October 2025 is not a cyclical downtrend, but another extended consolidation phase, similar to those we saw from March to October 2023 and again from March to October 2024.

A true Bitcoin cyclical downtrend usually involves a year-long drawdown of 50% to 85%, punctuated only by brief relief rallies. That is clearly not what we’ve seen. Instead, we had a sharp correction lasting about six weeks, after which Bitcoin bottomed and didn't do much.

When that bottom occurred, our MLDP Z-Score reached an extreme reading near –3 standard deviations. At that point, it became clear to us that we were at least dealing with a strong temporary bottom.

In the weeks that followed, price attempted to revisit those lows but consistently failed, forming higher lows after higher lows. That behavior, in itself, was constructive and increasingly bullish.

The next leg up

Although we believe Bitcoin is consolidating rather than entering a prolonged downtrend, I don’t think it is yet ready for a truly explosive rally. Lower interest rates should help Bitcoin in 2026, but beneath the surface we are not yet seeing the usual signals that precede major upside expansions.

Yes, on a very short-term basis, Bitcoin has just broken out. After consolidating in an ascending triangle—marked by consistently higher lows—it finally broke above resistance around $94,000 last week. Price then moved quickly toward $98,000 and is currently retesting that former resistance as support.

This move, if it hold, should be enough to push Bitcoin back above $100,000, potentially toward $110,000. But for a sustained breakout to new highs, several things still need to change.

Supply dynamics

The last two extended Bitcoin consolidations ended when the number of newcomers entering the network began to rise, while hodlers stopped selling

In that environment of 2023, SOPR (the indicator that triggered our exit signal in August 2025), was also trending meaningfully higher.

None of these conditions are fully in place today.

The number of newly created addresses with a non-zero balance, while significantly higher than the lows reached in June 2024, is currently sitting near levels that marked local lows twice over the past two years.

SOPR has only recently begun to trend slightly higher after a prolonged decline.

This is encouraging, but still far too early to call for an explosive move.

More positively, the amount of Bitcoin that hasn’t moved for an extended period has finally started to increase meaningfully since bottoming in early December.

At this stage of Bitcoin’s maturity, it is normal—and healthy—to see long-term holders selling into strength during uptrends. What we really want to see during corrections is holders stopping their selling. That behavior tends to put a floor under price and gradually applies upward pressure.

In both 2023 and 2024, a clearly bullish configuration across these indicators emerged roughly two weeks before Bitcoin’s next explosive move. The current setup, while improving, tells a different story: Bitcoin can continue to grind higher, but its next true breakout likely still needs a bit more time.

For us, that’s perfectly fine. We are still massively up on our Bitcoin position, having entered at the end of 2022. Bitcoin doesn’t need to be our MVP every single year for our thesis to remain valid. Our conviction in the asset remains strong, and we believe that in time—sooner rather than later—it will reach our target.

That said, one implication of our view that the traditional four-year cycle no longer strictly applies is that Bitcoin’s price dynamics could evolve very differently over time. In particular, it opens the door to the possibility of smaller, more irregular “mini-bear” phases. I’m fairly convinced that if the broader equity market were to experience a sharp correction, Bitcoin would likely follow. In such a scenario, it could take even longer before new highs are reached.

If that scenario were to materialize and align with one of our hedge signals, we might consider exiting part of the position in order to re-enter at lower levels. But at this stage, that remains highly speculative. Until such a scenario clearly plays out, we will simply continue to hold Bitcoin. We are positioned for the long game, and the opposite outcome likely has at least as much probability. Moreover, with MSCI deciding not to remove MicroStrategy from its indexes—a move that could have created selling pressure on MSTR and potentially forced Bitcoin sales—one major source of uncertainty around Bitcoin has now been removed.

If market breadth expands meaningfully this year, there is a strong chance Bitcoin will follow. Should that happen, we would begin trimming the position roughly 20%–25% under our actualized end-of-cycle target. When that time comes, we will publish an update revising our targets, because even though Bitcoin’s price has gone largely sideways over the past few months, the underlying fundamentals have continued to evolve. For example, Bitcoin’s realized value—the average price paid across all coins—was around $50,000 in mid-July, when we published our initial top estimate. Today, that same metric is closer to $56,000. That kind of shift will naturally reframe our estimates when Bitcoin resumes its upward move.

WU plan for 2026

We now arrive at the part of this traditional post that, this year, took me more time to write than usual—the section where I talk about our plan for the year ahead. But before I switch into full project-manager mode and start announcing colors and roadmaps, I want to begin with something more personal.

As AI began to pick up steam at the end of 2022, I was a university professor working in that field, in a city where AI generates a lot of noise. One of the three founders of deep learning is a professor at the University of Montreal and chose to remain here, which led governments and companies to invest heavily in the ecosystem. In that context, I was approached several times to “jump ship” and join other companies—sometimes with compensation packages far larger than what I had as a professor. Each time, I refused the call.

As I mentioned earlier, 2025 marked a turning point for Physical AI, with massive capital flowing into robotics and embodied intelligence. I created my LinkedIn account maybe twelve years ago and never really updated it. There’s very little on it—it still lists me as a university professor and co-founder of Robotiq. Robotics is not yet a massive industry, but within that space, Robotiq has become fairly well known. I suspect that seeing someone listed as a founder of one of the most successful companies in the field, yet not actively involved day to day as a manager, engineer, or scientist, sparked a lot of curiosity among headhunters.

As a result, I received this year an impressive number of messages for CTO-type roles and similar positions. One of them was particularly a big company—big enough to be genuinely flattering. But again, I refused the call.

Then, in November, my phone rang. This time, it was the CEO of Robotiq, asking me to come back. The request came with a clear vision: robotics was about to become far more technology-centric, with sensor technology and intelligence emerging as core challenges. This time, I accepted the call.

So here I am now, back at Robotiq as CTO-AI, with that objective in mind. One of the first building blocks of this vision was to spin into a product the tactile sensing platform my research team developed at the university between 2011 and 2017. Human grasping ability relies on an incredibly complex and rich sensorimotor system, and at the time, there was simply no equivalent available for robots. So we built one.

We produced about thirty prototypes and sent them to AI and control research teams around the world to explore what could be done if we added touch sensing to robots. You can learn more about that old initiative here. Today, I have the opportunity to turn that work into an actual product—one that we will be releasing in roughly three weeks.

One of the things I always disliked as a university professor was seeing great projects end up on a shelf once the funding cycle ended or students graduated. Having the chance to see all that work evolve into a real sense of touch—something AI models will use to learn how to interact with the physical world—is far more appealing than any of the other offers I received. And in a very strange twist of fate, I ended up rejoining this journey with WealthUmbrella co-founder Jennifer Kwiatkowski, who had joined Robotiq earlier last year as a Physical AI specialist, with a strong focus on intelligent grasping and manipulation.

This is only one of the projects I’m currently involved in; others remain confidential. Overall, coming back to Robotiq genuinely feels like the best place I could imagine putting my brain to work to help push AI into the physical world. I guess I’m now old enough to be at that point in life where you start wondering whether you’ll leave something meaningful behind when all of this is said and done.

This decision forced me to step back from my role as a university professor. When I made that decision last fall, it was already too late on the teaching side—students had already enrolled in my two winter courses. I also still have two PhD students and one master’s student under my supervision. As a result, I’m currently in a transition phase. That transition, which began in the first week of January, has been exhausting, but things are finally starting to normalize.

I became a university professor at a very young age compared to what is more typical. It has been a wonderful journey, and for many years it genuinely thrilled me. That said, it was never all candy. One of the less visible aspects of being a professor is the sheer diversity of tasks involved. From the outside, people mostly see teaching and research, but behind the scenes there is much more.

You have to write grants—many of which get rejected. When they are accepted, they come with a heavy administrative burden. You manage lab payroll, purchase consumables and equipment, handle and partially write patent applications, write papers, give conferences, sit on thesis committees, and participate in various academic boards. My grandfather, who spent part of his life as a university professor, warned me on day one of my career that this lack of focus was the main enemy—something that could slowly erode the love of academic research. In fact, although he continued working in industry until the age of 82 (in software simulation), he left academia much earlier for precisely that reason.

I was lucky early on with funding, and my lab grew very quickly—probably too quickly—around 2012. At its peak, I had around 25 graduate students, one or two postdocs, two research engineers, and one coordinator (Nathalie, who later joined me at WealthUmbrella).

(here with few of my students. I am the one hugging the robot)

During those intense years, almost all of my time was consumed by administration, grant writing, paper corrections, and weekly meetings with students. The hands-on research—the part I truly loved—was mostly being done by my research engineers and students. I started to find that frustrating and began to better understand what my grandfather had warned me about. To compensate, I tried to carve out one week every summer when I could actually sit down and do real technical work again. It was in that context that I started working on WealthUmbrella—a way for me to become an engineer again and keep my hands deep in technical problems.

Over time, I realized just how much I had missed that feeling. That’s why I gradually invested more and more time into WealthUmbrella. In 2025, I put an enormous amount of engineering effort into that project, which is also why we released so many new features. I had an incredible amount of fun coding TuneMap and the Downtrend Exhaustion Signals. I sometimes stayed awake very late—not out of urgency, but out of pure passion.

That same energy carried over into Pand-AI. Building it with Zackary was genuinely exciting. He is just as passionate about engineering as I am. You should see him when he starts digging into a project he loves—when he’s in that mode, I’m not even sure his wife can get his attention. Anyway, all of this reinforced a realization for me: maybe after 17 years as a university professor, it was time for a change. Maybe I preferred being part of the research again rather than mostly administering it. That was another key reason why coming back to Robotiq felt like the right move.

I’m sharing this story not to suggest that this is the end of my involvement with WealthUmbrella—quite the opposite—but to help you understand where my head is at right now. Naturally, this change will have an impact on WU, but not necessarily in a negative way. I will continue to be fully committed to our market updates. It might occasionally happen that you see an update written by Zack—for example, if markets crash in the middle of a product release—but my commitment to analyzing markets and writing for WU remains unchanged. After all, over the last few years I was already involved with Robotiq to some extent while simultaneously investing time in WealthUmbrella and being an active university professor. Losing one hat to reinforce another could actually be an improvement in terms of time management.

Where the impact will be more noticeable is likely on product development. Don’t expect the same pace of new feature releases as this past year. That said, this was already the plan even before I decided to step back into my robotics role.

I’ve had the vision for WU Advanced almost since day one of WealthUmbrella, and this year we pushed hard to finally materialize that vision. On top of that, we built Pand-AI, which was something I personally felt I needed as an investor. That sprint added a lot of new tools, and understandably, some members felt a bit lost. As a result, our focus this year will primarily be on helping people better understand everything that already exists on the platform. When we launch new indicators or products, we usually publish posts to explain them, and many of you who’ve been with us for years were able to digest that content gradually. For newer members, however, this now represents a lot of information to absorb all at once.

One of the first initiatives in that direction will be an evolution of the small “i” icons in our DataHub. Instead of a one-sentence tooltip, they will now open a more detailed—but still concise—explanation. The backend work for this is already completed, and it should be released very soon. It will look something like this:

We’re also in discussions with one of our members to create a Udacity-style course around our data-driven approach. The goal would be to explain the specific role of each signal and how to use them effectively in an investment process.

Along those lines, we may also reach the end-game we initially envisioned for Pand-AI: connecting it directly to our market signals for our S&P 500 members, so you can generate a proper market update whenever you feel the need. This remains uncertain, as there are still technical challenges involved, but it is certainly possible that we accomplish this in the second half of the year.

What is certain is that we plan this year to modernize our hedge algorithm. Not because it doesn’t work—on the contrary, it continues to perform in line with its historical statistics and is often better than me at assessing what’s coming (February 2025 being a good example). That said, aside from a minor adjustment in October 2023 due to Nasdaq discontinuing some data, we’ve essentially been running the same version for four years—a version that was originally designed five years ago and backtested using data going back to 2019.

After some internal, and frankly philosophical, discussions, we believe it’s time to incorporate more of that history into a new iteration. Markets evolve, and our signals need to evolve with them. The challenge is finding the right balance: preserving enough untouched history for proper validation and to avoid overfitting, while still leveraging more recent data. Our current plan is to use the 2019–2023 period to build a new version, which will likely also integrate additional new signals.

I consider our Downtrend Exhaustion signal to be the most accomplished work I’ve done on this project in term of signal processing. That’s probably normal—I had four more years of experience when I built it—but I also invested an enormous amount of time exploring hundreds of datasets. I distinctly remember feeling, once it was finished, that I had exhausted every reasonable option.

I was genuinely amazed to see how well it performed live in January 2025, again in April 2025, and more recently at the bottom of the last small correction. I believe the underlying data used by some of these signals is extremely powerful, and I think there is room to repurpose part of that work to improve our hedge signal. That’s likely where I’ll spend most of my personal technical time this year, as the hedge signal is the sun around which the rest of our tools orbit.

So thank you for being with us through an exciting 2025. I don’t expect 2026 to be an easy market, which probably means you’ll be seeing me write quite a bit. That said, there should also be some good moments along the way—ones that should ultimately push our investment portfolios higher (maybe just not tomorrow morning, given what happened over Greenland this weekend).

I wish you all a great 2026 investing year.

Vincent