LITLighterNeutralLighter

Today on

Why Robot Hands Are So Hard to Get Right

With this packed earnings period we’ve spent so much time going through company results. Let’s take a step back for a bit and dig into some humanoid research. I want to spend this one on the hands. In the earlier humanoid article we went through what could get these robots from a few units being tested to hundreds of them actually working. And a lot of it comes back to something that seems pretty basic. Can the robot pick something up and do the job with it? Keep doing it, all day, with someone coming over to help only occasionally? The hands are so intricate. Think about it. Pick up a pen and turn it around in your hand. You probably moved a few fingers, loosened your grip in one spot and tightened it somewhere else. It takes almost no thought. Now try to describe all of that to someone building a hand from scratch. Where do the motors go? How does it feel the pen starting to slide? How hard should it squeeze? There is a lot going on in this tiny space, and the decisions start affecting each other pretty quickly. Some of the companies we’ve been talking about are already working on pieces of this. ALGM, NVEC, VPG. We have named integrations, customer orders and sensor development to look at. First, look at this. A Shadow hand holding a light bulb, photographed in 2007. All that hardware coming through the wrist and these fingers wrapped around thin glass. People have been at this for a long time! Just getting the fingers to close around something took a lot of work. Then you start asking the hand to use it. Picking it up, then doing something with it Back to the pen. Picking it up gets it off the table. To write with it, you have to settle it between your fingers, turn the point toward the paper and hold it there as you move. The object moves around inside your grip. Some fingers let it slide while others keep it from falling. Your thumb comes across from another direction and helps hold the whole thing together. ‘ Dexterous manipulation’. For plenty of factory jobs, the part shows up in the same spot and just needs to be moved somewhere else. Look at Baxter’s two-finger gripper below. Close around the part, move it, release it. If you can do the whole job that way, you have fewer moving pieces to buy and keep working. A customer can get quite a lot out of a setup like this. But say that same robot is going to open a drawer, take a tool out and then use the tool. Now the grip changes as the job changes. A hand with more movement has more ways to get through that sequence. And we already have drawers, handles, tools and workstations everywhere that were built around people using them. Getting a robot to work with what is already there would save a lot of redesigning around the robot. I think it’s quite easy to see the appeal, but you can also see how quickly the hand itself gets complicated. How the fingers move Let’s start with how a finger bends. Some of the muscles moving your own fingers sit back in your forearm, and tendons carry the force through the wrist. Follow them in the drawing below. They have quite a distance to travel before they reach the finger. So you can put the motors farther back and run tendons out to the fingers, much like our own hands. Lighter fingers, more room at the tips. Makes sense. But now you have this tendon running from the motor out to the joint and it stretches a bit. It rubs against the guides. The tension changes as the finger bends, and you keep doing that over and over. Eventually it wears and someone has to replace it. Then get the tension right again. Even the decision to move a motor farther back follows you all the way through to the person fixing the hand later. Or you can fit more of the machinery into the hand itself, using gears and rigid links to move the joints. We’ve talked about actuators before. For an electric hand, think of the motor and the mechanism carrying its movement out to the joint. Put that assembly close to the finger and the finger carries more hardware. Move it farther back and you have to get the movement out there somehow. There are different movements to account for too. Curl your index finger. Then spread it sideways. Your thumb can come across your palm and meet the other fingers, which is especially helpful for holding little things. These independent ways of moving are called degrees of freedom (DoF). A hand might have a big DoF number, but the placement of those movements tells you a lot more about what it can do. Having a thumb that can get around an object changes the grip available to the whole hand. You can also use a single motor to bend several joints together. Pull a tendon, the finger closes, and when one section hits the object the other joints can keep wrapping around it. That’s called underactuation. Fewer actuators than degrees of freedom. It’s a way to get a useful grasp while saving some hardware. For moving the pen around inside the hand, separately controlled joints give you more options. Then the software has to coordinate those options. Here’s an example from a 2021 research paper . The researchers built the motors, links and sensing into the hand, which they call ILDA. You can see the whole assembly on its own and attached to an arm. And here is one finger opened up. Look at the little links connecting the joints. All of that has to fit inside something narrow enough to work alongside the other fingers. Want a stronger grip? A bigger motor can help, and gears can turn fast motor rotation into slower, stronger movement. Then you have the weight of that motor, the heat it makes and the friction in the gears. There’s a little play in the mechanism too. Add the bearings, wires and sensors, then put a cover around everything. Even if you get it all to fit, the arm still has to lift it. So a change that looks pretty straightforward at the motor can end up affecting the whole robot. MinebeaMitsumi showed a hand prototype at CES 2026 with 85 miniature and small bearings and 11 actuators. It developed the drive unit with Harmonic Drive Systems. 85 bearings in one hand. That makes the mechanical side of this easier to picture. Every joint has to move smoothly in a small space, over and over. This is a prototype, so it gives us an example of the hardware being developed. As these designs get into production, the particular parts that survive all that testing are the ones a customer would keep ordering. Even the stiffness is something you can design around. Let a finger give a little when it bumps into an object and it can absorb some of that unexpected contact. ‘Compliance’. Researchers at the German Aerospace Center have a tendon-driven hand with adjustable stiffness . It can be firmer for one task and more forgiving for another. Think about first reaching for the bulb, then holding it while the arm moves. Those are different demands on the same fingers. Getting the finger to the right place Read more

NeutralAAOIApplied Optoelectronics, Inc.Applied Optoelectronics, Inc.IRENIREN LIMITEDIREN LIMITEDIWMRussell 2000Russell 2000LITELumentum Holdings Inc.Lumentum Holdings Inc.MRVLMarvellMarvellQQQNasdaqNasdaq

2d ago on

The AI Selloff, Anthropic’s Growth and Where I’m Looking Next

This week gave us a lot to work through. We got a closer look at Anthropic’s business, a wave of earnings across AI infrastructure and Physical AI, and a selloff that brought some of the companies I’ve wanted to own back into view. I’ve spent the weekend going through all of it and narrowing down where I want to put my attention next. Some of these are businesses I already like and want to buy at the right price. Others are names outside the AI buildout where the charts are starting to look interesting. Let’s get it started. TLDR + Contents This week’s research: A recap of my earnings and data-center articles and the biggest takeaways. The selloff: Where the selling hit hardest and how other sectors held up. AI economics: What Anthropic’s latest filing tells us about adoption, spending and profitability. The rotation: Where I’m seeing early signs of strength outside the AI trade. Power and cooling: A company I think is well positioned to meet the changing demands on data-center developers, and the price where I’d be more interested. Outside tech: Three companies I’m considering for a potential rotation, with the earnings, valuations and charts behind each setup. Software: A speculative recovery candidate with earnings approaching. Physical AI: Two companies with minimal data center exposure. I want to build larger positions in as machines become more capable. Networking: Two companies I think are well positioned for continued AI infrastructure spending. 24/5 trading: Two companies positioned to benefit as stock trading expands beyond regular market hours. This week’s research The earnings targets and hyperscaler spending questions heading into the week. I also dug into OpenAI’s reported fundraising plans and the newer spending data from Ramp and OpenRouter. The question was how much of the infrastructure already committed could be supported by customer revenue, and where financing delays could push supplier shipments further out. LITE, FN, AEHR, TSEM, LSCC and ON, plus the MRVL–LITE deal. Power, water and local approval requirements, including the different implications for operating facilities and projects awaiting permits. Texas represented roughly 32% of the tracked U.S. data-center capacity under construction in the Baxtel dataset, so the timing of new rules matters. I explored how tighter supply could increase the value of existing capacity while also leaving providers unable to serve the next customer. The week’s broader earnings coverage across networking and Physical AI. This brings eight company updates into one piece, so readers following several of these names can revisit them together. AAOI’s ramp timing and the other company updates, including INDI’s production progress and financing considerations. Community demands, the construction pipeline and a closer look at IREN’s funding and delivery schedules. I spent more time on the interviews and what developers are being asked to deliver around power, water and local approval. The selloff Well, that was a rough week. The S&P 500 finished down 2.7%, the Nasdaq fell 3.4%, and the Dow was down 0.9%. Small caps also struggled, with the Russell 2000 down 2.1%. Friday’s bounce helped, but it didn’t undo the damage from earlier in the week. And depending on what you owned, it probably felt a lot worse than those index numbers suggest. Semis in particular have taken a beating. Looking at this chart, I’m having a hard time getting too excited about one green day. There have been plenty of bounces along the way down. For now, I want to see support hold and some higher lows start to form before getting too aggressive. I’ll get into the levels I’m watching on the indexes later and how they’re influencing my plan for this week. Anthropic Since I wrote about Anthropic’s filing, we got another look at how quickly the business is growing. The Wall Street Journal reported that its annualized revenue run rate topped $80 billion at the end of October, citing a person familiar with the business. Anthropic generated $17 billion in revenue in Q3, but the cost of serving all that demand increased. For every dollar it spent on compute and infrastructure, it generated $1.72 in revenue. That was down from $2.08 in Q2. A big part of that was the roughly $5 billion it spent on short-term compute commitments. That capacity is more expensive than what it can secure through multiyear agreements. So demand is growing fast enough that Anthropic is paying up for capacity it can use now. As more contracted infrastructure comes online, it has an opportunity to move workloads onto cheaper capacity. That’s good for its margins and gives infrastructure providers a clearer picture of the demand they’re building to serve. Adjusted operating profit rose to $2.5 billion in Q3, but the margin fell from 19% to 15%. On a GAAP basis, it reported a $567 million operating loss, including $1.07 billion in stock compensation and $1.96 billion in noncash donation expense. We’ve already talked about the growth in its customer base and the drop in concentration. One major thing keeping an eye on now is how quickly Anthropic can bring the cost of serving those customers down as demand keeps growing. The rotation I’ve been talking for weeks about broadening my exposure to other industries. My concern now is that a lot of the strongest non-tech stocks have already had big runs. If money rotates back into technology, I could end up prioritizing new additions away from a trade as it gets cheaper and buying another after a big move. Last week showed how uneven the market has become. Technology fell 8.23%, while communication services, real estate, consumer discretionary, financials and consumer staples finished higher. That’s why I’ve spent so much time looking for stocks that are still building bases or starting early reversals. I want businesses I’m comfortable owning, with valuations that make sense and charts that give me a reasonable place to start. I’m not looking to abandon AI obviously. I still have so much exposure. I want to keep the companies I think are best positioned while adding businesses with different sources of demand. Below, I’ll get into the companies I’m considering and how I’m approaching each one, from power and cooling to non-tech, Physical AI, networking and longer trading hours. Financial Disclaimer This article is for informational and educational purposes only and is not personalized investment advice or a recommendation to buy or sell any security. The views expressed are my own and may change without notice. I may own positions in companies discussed and may buy or sell shares at any time. Investing involves risk, including the possible loss of principal. Do your own research and consider your personal financial situation before making investment decisions. Read more

IRENIREN LIMITEDNeutralIREN LIMITED

3d ago on

The Data Center Backlash Is Here. What Happens Next?

If you own AI infrastructure stocks, these developments are something worth tracking. A lot of the growth we’re expecting obviously depends on data centers getting built and connected to power. If approvals take longer, projects cost more or companies have to find another location, that can change when the revenue shows up. But how much of the buildout is actually at risk? In my article after the Texas election , I covered the initial implications. Since then, I’ve been going through the interviews, local decisions and company calls to get a better sense of what could change. There’s still a lot being built of course. Some developers are already addressing the concerns. Others could face a much harder approval process. I think we need to understand those differences so we can sift through the different downstream impacts. That means looking at what communities are asking for, what companies are agreeing to, and how far these projects have actually progressed. Let’s get into it. TLDR + Table Of Contents Local tradeoffs . Tax revenue can be enormous. That doesn’t resolve the complaints of people living beside a facility. Who pays? Utilities need protection if developers use less power than promised. The agreements determine who carries that risk. Community deals. Developers are offering water infrastructure and bill credits. Some benefits are funded, others still need approval. New rules. Tax conditions, temporary pauses and possible prohibitions have different consequences for construction. IREN. Real delivery progress and customer funding, alongside timing changes and contractual obligations. The buildout. A substantial pipeline is underway. Earlier spending is one possible response to restrictions, but delays are another. Read more

MRVLMarvellBullishMarvell

4d ago on

$MRVL is going to be such a strong company if this goes through Definitely one that will be at the top of the list for anyone investing in networking No position for me yet https://t.co/lqlUZk3fTk

COHRCoherent Corp.NeutralCoherent Corp.

5d ago on

Nice to see $COHR with more diversification in their 6-in substrate suppliers https://t.co/7QULZK57gU