What We’re Reading (Week Ending 12 April 2026) - 12 Apr 2026
Reading helps us learn about the world and it is a really important aspect of investing. The late Charlie Munger even went so far as to say that “I don’t think you can get to be a really good investor over a broad range without doing a massive amount of reading.” We (the co-founders of Compounder Fund) read widely across a range of topics, including investing, business, technology, and the world in general. We want to regularly share the best articles we’ve come across recently. Here they are (for the week ending 12 April 2026):
1. America’s AI Build-Out Hinges on Chinese Electrical Parts – Emily Forgash and Akshat Rathi
Almost half of the US data centers planned for this year are expected to be delayed or canceled. One big reason is the shortage of electrical equipment, such as transformers, switchgear and batteries. They are needed not just for powering AI, but also for building out the grid that is seeing increased consumption from electric cars and heat pumps. US manufacturing capacity for these devices cannot keep up with demand, and the scarcity has caused data center builders to rely on imports…
…Data centers consuming as much as 12 gigawatts of power are supposed to come online in 2026 in the US, according to analysts at market intelligence firm Sightline Climate, who will be releasing a new report in the coming weeks. However, only a third of that is currently under construction, Sightline estimates…
…Electrical infrastructure adds up to less than 10% of the total cost of the data center, but it’s impossible to build the operation without it. “If one piece of your supply chain is delayed, then your whole project can’t deliver,” says Andrew Likens, Crusoe’s energy and infrastructure lead. “It is a pretty wild puzzle at the moment.”…
…Though few companies are eager to talk about it, the US has been outsourcing its manufacturing to other countries, primarily China, for decades. That has contributed to a significant shortage of electrical components in the US, says WoodMac’s Boucher…
…While most of the US’s transformers come from Canada, Mexico and South Korea, US utilities imported more than 8,000 high-power transformers in 2025 through October from China, up from fewer than 1,500 imported in all of 2022, estimates WoodMac’s Boucher. This build-out “is going to be highly dependent on the import market,” he says.
Once transformers lower the voltage of electricity so it can be used in data centers, it then needs to be distributed across the data center safely. That’s done through switchgear, which includes circuit breakers and fuses. There too, data center developers are seeing delivery delays – though not as extreme as the timelines for transformers.
Equinix Inc.’s solution is to commit at least $350 million to support Hanley Energy’s new manufacturing facility in Ireland, which will make switchgear and other data center components. Equinix expects to achieve 10% to 15% faster lead times as a result.
Crusoe’s answer to that shortfall has been to pre-order lots of the equipment. That means spending many millions of dollars on supplies before the company even knows it has an order to fill, but it’s proved a winning strategy. Recently, Crusoe also began manufacturing their own switchgear…
…The share of US imports of transformers and switchgear from China has declined steadily in recent years – although for specific types of equipment that share is still hovering around 30%. The Chinese share of battery import volumes remains stubbornly above 40%.
China dominates the supply of electrical equipment because it controls so many parts of the supply chain, from materials to processing to manufacturing, and the gulf between China and the US is set to widen. In its new five-year plan, the Asian giant revealed last month that it will double down on building out its grid with renewables, while the Trump administration has dismantled policies to deploy solar and wind power.
2. “Founder Mode” Complacency – Abdullah Al-Rezwan
When DeepMind was plotting to extricate themselves from Alphabet almost a decade ago, Pichai was prescient enough to foresee AI’s paramount importance in their core business…
…As these negotiations became more tense over time, all the big guns of Alphabet planned to meet to resolve the issue at hand. Alas, some big guns didn’t seem to appreciate what was at stake. From the book:
When the two sides met again, the conversation underscored the gulf between them. Hassabis and Suleyman argued that DeepMind did not fit under Google’s umbrella: Its mission was AGI, not consumer‑internet products. Pichai objected that AI was central to his vision for Google, and that he would not allow his scientific bench to be depleted. Hassabis had hoped that Larry Page would weigh in on his side and push the Alphabet plan to a conclusion. But Page showed up for the meeting two hours late, and Sergey Brin was even later. Their version of what later came to be known as “founder mode” was that they were nowhere to be found, disproving the Silicon Valley mantra that founders deserve the right to control their companies indefinitely. With Page and Brin effectively checked out, Pichai was the man DeepMind had to deal with.
I have been thinking about the aforementioned excerpt for the last couple of days. If you glanced at my portfolio, it’s not difficult to see that I drank my fair share of kool-aid of “founder mode”. Perhaps fittingly the “founder mode” propaganda originated from a founder himself: Brian Chesky. The more I ruminated over “founder mode”, the more I came to the conclusion that there is a glaring missing aspect in “founder mode” mantra: Complacency.
It is telling that Chesky proudly recalls every chance he gets about how he figured out during Covid that Airbnb doesn’t need to do search advertising; as an investor I was actually a bit alarmed that he was running Airbnb pre-pandemic without paying close attention whether his advertising dollars was being deployed with appropriate ROAS guardrail. I can guarantee you that despite operating in “Manager Mode”, Glenn Fogel at Booking was looking at advertising ROI with a microscope and he certainly didn’t need a global pandemic to remind him how to deploy his precious advertising dollars at Booking.
3. A token is not a fixed unit of cost – Anjali Shrivastava
We only consider token count as the static linear meter because we inherited the logic from inference APIs. But, a token does not represent a fixed unit of work.
This is obvious to anyone who works in inference, but if you’re used to calculating compute budgets based on linear API rates, it takes a second to sink in.
The intuition is grounded in the autoregressive nature of the transformer: Attention is quadratic with respect to current context size…
…In layman’s terms, the language model is looking at every previous token in the context window before generating a new token, which means inference APIs are linearly pricing fresh tokens whose compute cost scales quadratically.
The scaling law for compute is likely not purely quadratic, given optimizations like caching and compacting context. But no matter what, the underlying compute cost per token grows with context length. The Nth token in a conversation is an order of magnitude more expensive than the first.
There’s signs that per-token pricing breaks down at scale: both Anthropic and Google charge different rates based on prompt length…
…Traditional SaaS has variable costs too (like hosting, customer support and third-party service costs). But these costs follow the law of large numbers, and are normally distributed at scale. You can set a single subscription price that covers this average cost, plus a comfortable margin to absorb tail risk.
In the case of AI software, it is likely that these variable costs are fat tailed. The law of large numbers assumes finite mean and i.i.d. samples, but AI software has at least one dimension with infinite first moment and non-stationary tails. The sample mean keeps wandering instead of converging…
…Margin collapse is the first and most obvious symptom of the problem. Cursor’s repricing exposed poor margins, and we also learned that Replit’s margins are volatile. And there is ample evidence that Anthropic is losing money on its subscriptions.
Each layer of the aggregate cost curve is highly variable, and the more you scale, the higher the probability that these tail risks can compound…
…Subscriptions misprice intelligence, and much of the industry recognizes this, but now we can rigorously explain why.
Traditional SaaS pricing mirrors the physics of stable software, but AI introduces high variance that breaks each of these laws…
…High variance in costs necessarily constrains demand; today, the constraints are reactive.
To safely cushion from unbounded costs, a business model must price in the variance or be well above the true cost on average. Ideally by anchoring price to value delivered instead of token cost; but value delivered also happens to be highly variable and subjective. At the same time, there’s structure to value: reliability, relevance, actionability.
The key insight is that margin squeeze and resource misallocation are two sides of the same problem. Solving one side of the equation should solve the other. If you can measure the value delivered, you can price that instead of raw compute. And if you can price outcomes in terms of value delivered, you can budget the exact amount of compute and data that maximizes profit on each task.
So the layer that owns the meter also decides how much compute and data to deploy and keeps the spread between cost and price. Today that meter sits inside the model; tomorrow it could sit inside an orchestrator that plans the whole workflow.
4. Why You Should Wait Out AI’s Super-Spending False Start – Merryn Somerset Webb
The second part, the data on which all LLMs are trained, is not. Its supply is limited. Up to ChatGPT4, the internet provided enough data for each new iteration to be better. But that version was completed a few years ago, trained on the lot. There is little more for new models to train on.
The data on the internet might have expanded over the last few years, but not in a particularly helpful way. Much of it has been produced by other LLMs: train your new model on that and you might end up degrading it. Why? Because LLMs are horribly prone to errors (confabulations or hallucinations), which means they can’t give us what we most need from them: accuracy.
An LLM is not a continuous learning machine. Its knowledge stops with its training. It also isn’t deterministic (like, say, a calculator), says AI expert Janusz Marecki (who I interviewed for a podcast this week). It knows nothing with certainty. It simply “rolls the dice” on what the next word in a series should be, giving you its best guess. The answer you get is an approximation, not a series of facts. Worse, the more complicated the task in hand, the more the errors compound. Possibly even worse, the LLM can’t tell you how likely it is that there are errors. How would it know?
These problems aren’t going to go away. They are irredeemable systemic flaws in the product.
5. Switzerland – Europe’s overlooked activist opportunity – Swen Lorenz
Switzerland is famously conservative and generally averse to outsiders telling it what to do.
This is also reflected in its corporate landscape.
Even though the country is broadly open to foreign investment, there have long been numerous mechanisms allowing companies to keep outside influence under tight control.
Some Swiss companies require shareholders to be registered by name, with board approval needed for new registrations. This has led to cases where outsiders were refused registration – and “outsiders” can even include Swiss citizens from a different region.
Other companies cap voting rights per shareholder or maintain super-voting shares that remain tightly held by local incumbents…
…The 2023 reform of Swiss corporate law wasn’t widely noticed, not least because attention was focused on events in Ukraine and the aftermath of the pandemic.
Until then, a shareholder needed to represent 10% of share capital to add an agenda item for a vote at the annual general meeting.
For publicly listed companies, this threshold has now been reduced to just 0.5% – a far more attainable level.
Similarly, a shareholder with 5% can now requisition a shareholders’ meeting, compared to 10% previously.
Just as importantly, the broader acceptance of active shareholders has evolved…
…Finanz und Wirtschaft, Switzerland’s leading German-language business daily, carries significant influence among corporate executives. In an article published on 18 September 2025, the paper noted how “activist investors are transforming from bogeyman to catalyst”…
…Patrick Fournier is an active investor based in Zug. We met several years ago at his family home to discuss our shared interest in frontier markets.
Today, his focus has shifted closer to home.
He allowed me to share the following:
“We have progressively sold all our portfolio of foreign shares and are now focusing on Swiss small & mid cap. We see huge value opportunities on this segment. We intend to become a little ‘activist’ as it is now possible with only 0.5% of capital in a listed company (far lower than the previous 10%) to add some proposition at the agenda of the annual general meeting of shareholders. This will wake up the Board of several companies, including regarding the dividend (payout) policy. As a result, we are in front of a ‘rerating’ (multiple expansion) of this segment.”…
…BVZ held its annual general meeting on 8 April 2026, and the results were telling.
Some 287 shareholders attended, representing 110,328 out of 197,278 shares outstanding (with one shareholder alone holding 56,000 shares). Alarick’s proposal to increase the dividend from CHF 18 to CHF 50 received 14.5% support and was rejected by 83.8%. As a result, the board’s proposal to raise the dividend from CHF 16 to CHF 18 was approved. With earnings per share of CHF 151, this implies a payout ratio below 20%. The proposal to initiate a share buyback programme received 16.67% support and was rejected by 82%, and therefore did not pass.
What may sound like a defeat is, in fact, the equivalent of an earthquake. In Switzerland’s highly consensus-driven corporate culture, such a level of shareholder dissent represents a clear wake-up call for management.
The market agreed. On the day of the meeting, the share price closed at an all-time high of CHF 1,550, up 67% over the past 12 months.
As the recent share price performance suggests, even raising one’s voice in a constructive manner can create value for shareholders in Swiss companies.
Disclaimer: None of the information or analysis presented is intended to form the basis for any offer or recommendation. We currently have a vested interest in Alphabet. Holdings are subject to change at any time.
