TSLA: Good News, Bad News, How to Win Big

30 Dec, 2025

In 1989, Richard Gabriel wrote 'Worse is Better', the argument that Unix and C won not by being elegant, but by shipping and iterating relentlessly. Early Tesla embodied "worse is better" in spirit: ship an imperfect car, fix it over the air, move faster than anyone else. Somewhere along the way, Tesla became what it beat.

I first heard about Tesla around 2012, when "208-265 miles of range" (depending on battery) sounded laughably short. It might be a science project, not a mass-market car. Then I read the first Master Plan, watched the early-adopter wave, and noticed a pattern: nobody could go back to gas, and I changed my mind.

Over the next decade, I went through the whole shebang: watched the stock rip, bought too early and sold too early, then built conviction as the product and execution proved real. In 2016, I rode in a friend's Model S, a friend who had trashed Tesla in college and now couldn't stop praising it. Then I bought my Model 3, which reinforced my conviction. I held TSLA through the rate hikes and pandemic and experienced, for the first time, capital gains that exceeded my W-2. Later, when the CEO drifted into Twitter and politics, I started trimming and eventually exited by 2023.

So here's where I landed: Tesla isn't a bad company, but it's valued as a self-driving + robotics company that is about to explode without the execution to match that.

The path to justifying that valuation is narrow, but it only exists if Tesla starts executing like a relentless machine again.

Good News: Tesla still has real moats

Tesla proved the "EV as a better product" thesis. People didn't buy Teslas to save the planet; they bought them because the product was simply better. The powertrain efficiency is real engineering. The software-first feel is real. Smooth animation that still lacks in some of the legacy automakers today. OTA updates rewired customer expectations. My college friend who once trashed Tesla ended up buying a Model S and did a 180: "This is what a modern car should be."

Tesla survived the hardest chapter. 2017–2019 was trench warfare: production hell, bankruptcy narratives, relentless bears calling it a fraud. Tesla got through it. It hit scale. It proved it could manufacture. That matters.

The company still has leverage that most automakers don't. Brand¹ and mindshare. Vertical integration that can move fast when focused. A huge fleet producing data. Software DNA that legacy automakers dreamed of having. Remember when VW couldn't ship the ID.3 because the software was a disaster?

Tesla still has the ingredients to win big. It's not a "no hope" company. It's a "your valuation assumes you'll be at your best" company.

Bad News: The bull thesis cannot continue long without a result

Tesla has earned a lot of goodwill given its track record in the “happy time” (2019-2022), but this is not without its limits. If Tesla continues to execute slowly or get distracted, without delivering the goods, its valuation will fall back to earth.

Autonomy is a scaling race, and Tesla is losing on compute

My original thesis was "Tesla is the Apple of cars." Software + hardware integration lead, then iterate on the product to close other gaps. Think Apple versus Nokia, Apple might not have the best camera or the SoC initially, but by iPhone 4, Apple has caught up on camera quality, and its chip reaches SOTA (state of the art) by iPhone 5S.

But on autonomy, where the future value lies, Tesla fell behind. I expected Tesla to lead, not chase. You can't win by chasing. After Waymo's commercial deployment and scaling, Tesla's "Self-Driving" beta looks half-baked, still requiring geofencing and human safety drivers while Waymo operates fully driverless. This totally debunked the theory that Tesla’s system can work everywhere all at once. What I suspect happened behind the scenes is just oversampling/over-weighting the training examples, or special finetuning in the geo-fenced zone. If Tesla trusted its own plan, it wouldn’t do geo-fencing.

The bears say Tesla handicapped itself by going camera-only. Some say it’s the L2 vs L4 architecture.

The bulls say Waymo's sensor suite is too expensive to scale.

I say neither. They are distractions. The fact is that compare to compute, lidars are not that expensive. Lidar runs around $10K, Waymo’s compute stack, call it 4x H100, might run $30-40K.

The sensor isn't the bottleneck, the compute is.

And if you choose camera-only, that increases the burden on compute and models. Not to mention that we might want to put VLMs (Vision-Language Models) in cars in the future. You don't get to pick hard mode and use an inferior chip at the same time (unless your competitors’ models are garbage, but I doubt that).

Here's Tesla's FSD hardware cadence vs Nvidia Drive’s using car delivery date:

Tesla FSD Hardware

Nvidia Drive

The cycle got longer. That's backwards.

Apple ships new silicon annually. Nvidia every two years. Granted that automotive chips have a longer validation process, it is still slow. Compare to Nvidia’s side hustle, Nvidia Drive, Tesla’s current hardware lags behind. If Tesla stuck with Nvidia hardware, we would have 500+ TOPS HW5 by mid-2025. I expect better from Tesla’s vertical integration. Tesla went from startup speed to bureaucratic speed while claiming to lead the autonomy race.

The compute gap is real. HW4 delivers a guesstimated 250 TOPS. Waymo’s vehicle can do at least 2,000+ TOPS. Even if Tesla has a better model, the gap is hard to close, at least to me. And I fully expect Waymo’s next generation will have even more on-board compute.

The deep learning lesson is simple: the fastest way to improve performance is scale. Tesla has the data advantage, billions of miles of real-world driving, which is also a reason I invested in Tesla in the first place. But data without compute is useless. You can distill models to run on weaker hardware, but there is a limit. Smaller models are always worse given the same architecture; you can’t cheat here.

And this isn't speculation, the compute constraint is not just theoretical; it also shows up in the product. Every step change in driving quality comes from scaling the model. Since FSDv12, there has been no significant improvement. I believe the latest v13 model is already reaching the limit of HW4. As for my Model S’ HW3, even Elon admitted that it is inadequate and its performance is much worse than HW4.

So if on-board compute is the bottleneck, why did Tesla stop relentlessly iterating on it? I don’t know the answer for sure.

HW3 was promising, Tesla's own inference chip, shipping in vehicles, in volume. The right move was to iterate more: HW4 in two years, HW5 two years after that. Instead, the cycle stretched to four years, and resources went elsewhere.

Where? My guess is Dojo, an in-house training supercomputer announced around 2019. But Dojo wasn't just ambitious. It was exotic. Not Nvidia-style tensor cores. Not Google TPU-style systolic arrays. A completely unproven architecture with custom packaging. I knew it was set up for failure watching the AI Day presentation. However, at the time, Tesla had a huge lead, and it wasn’t an issue from an investment perspective.

The proven path in AI silicon is to start with known designs and iterate. Nvidia built tensor cores on top of GPU architecture they understood. Google evolved TPUs over generations, starting with inference. Tesla tried to leapfrog everyone with an exotic design. What’s worse, the team spent six years doing it.

Even if Dojo had worked, it solved the wrong problem. Training compute can be bought from Nvidia, but less so the inference compute that ships with every car. That has to be cost-optimized. Tesla had the right focus with HW3, then lost it chasing a Wunderwaffe.

Optimus is not "FSD with arms"

I'm not saying humanoid robotics can't happen. I'm saying the timeline investors emotionally price in is detached from the hardest constraints.

Hand manipulation is unsolved. Deformable objects are exceptionally hard. Data collection for driving seems like a piece of cake compared to humanoid robots. We simply don’t have enough robots collecting data to train a competent model. Cars have economic value without AI. Humanoid robots, on the other hand, are useless without AI, so it's hard to justify building a large fleet. It essentially becomes a chicken-and-egg problem. "Millions of robots in factories soon" is, at best, a fairytale. Even Ilya Sutskever can’t feel the AGI anymore these days.

If Optimus works, it would be because Tesla invested consistently for years with a consistent roadmap, not because demos magically turned into mass deployment.

"Soon™"became a valuation problem

There was a period when Elon-time was forgivable because the stock was cheap relative to the ambition, and execution wins were frequent enough to reset credibility. I held through those years. The product kept shipping.

However, today what we have are real soon now technologies: both FSD and Optimus. When you're valued like the future is imminent, repeated "almost there" cycles become a growing discount rate.

Tesla the car company can still be good. Tesla the stock can still be expensive.

Quality stagnated instead of improving

The Apple analogy required Tesla to do what Apple did: excel at hardware-software integration, then close the remaining gaps relentlessly. iPhone cameras were mediocre in 2007. By 2015, they set the standard. iPhone’s chip wasn’t the best, but now it is.

Tesla was supposed to do the same: panel gaps, road noise, fit-and-finish. It didn't happen. My Model S had the same complaints in 2022 that reviewers noted in 2018. Tesla's quality stayed flat while Chinese competitors closed the software gap.

How to Win Big: Rebuild the iteration cadence

Despite everything above, I still think Tesla can win big. It just needs to stop behaving like it already did.

1) Return to a predictable hardware cadence

Tesla’s current refresh cycle feels like a traditional automaker. If Tesla wants to be valued like an AI company, it needs to act like one.

That means a regular hardware refresh cycle, which means HW5 in 2025, not 2027. A drumbeat, not a question mark. An upgrade path that doesn't strand customers for half a decade.

2) Close the quality gap

If Tesla wants Apple-like value capture, it needs Apple-like hardware quality and refinement, no more giant panel gaps, and no more creaking noise.

3) Position Optimus honestly

There's a difference between "we're investing because this could be enormous" and "this will drive the business soon."

Tesla can do the first and stay credible. Tesla cannot do the second repeatedly without paying a compounding trust penalty. Most importantly, invest consistently, but do not divert resources from FSDs. (I heard some rumors that FSD engineers were diverted to help with the robot orgy demo; that’s not good.)

4) Focus

This is the uncomfortable part: Tesla's execution quality is linked to organizational focus, and organizational focus is linked to leadership behavior.

Fewer narrative pivots. Fewer distractions. More shipping.

Tesla doesn't need a new playbook. It needs its old one and a leader focused on executing it. Maybe try sleeping on the data center floor.

What would change my mind?

This is my personal scoreboard:

1. Hardware cadence speeds up. The first thing is to get HW5 out the door.
2. Quality improves in boring, obvious ways.
3. Optimus is positioned as a long-term bet, not a near-term revenue.
4. Sustained focus from leadership.

If Tesla does those things, it can justify a premium valuation, not because the market is gullible, but because the company would be compounding again.

Closing

The assets are still there. The brand, the data, the manufacturing, the software DNA. Tesla could still win big, but only by iterating fast again. Right now, the stock is priced for a fairytale. If the cadence stays slow, the price will fall.