Recently, I read two posts on X discussing the economic value of AI. Both were written in response to the so-called “Kimi moment”, which raised concerns that highly capable Chinese open models could undermine the profitability of US frontier AI companies and eventually burst the US AI bubble.

The core argument is that the profitability of selling AI tokens is not the same as the broader economic value created by AI.

Open models, model distillation and lower-cost Chinese competitors may compress the margins of leading AI labs. This could weaken their existing business models and make it more difficult for them to fund future training runs through token sales alone.

However, cheaper AI would benefit consumers, enterprises, cloud providers and software companies. AI can be monetised through workplace automation, enterprise software, cloud services and specialised applications, rather than solely through the sale of inference tokens.

The US government therefore does not necessarily need to protect the business model of any particular AI lab. If selling tokens becomes less profitable, these companies will need to adapt. Frontier-model training could still be funded by hyperscalers, industry consortia, governments or companies that generate profits from AI elsewhere in the technology stack.

The main uncertainty is whether cheaper models would weaken the incentives to fund expensive frontier research. This concern matters most if one believes that AI development is highly path-dependent and that the first lab to achieve a major breakthrough could secure a lasting advantage. Otherwise, lower prices and stronger competition are likely to benefit the broader economy.

The two original posts were fairly technical and contained many abbreviations and specialised terms, making them difficult to follow, especially for readers who are less familiar with the subject. I therefore asked ChatGPT to turn their arguments into a more detailed and accessible write-up. You can read it below.

What Jon Stokes and Nic Carter Are Arguing About Open Models, AI Economics, and the Future of Frontier Labs

Jon Stokes’s argument: OpenAI’s business model is not the same thing as the future of AI

Jon Stokes begins by trying to reconstruct the strongest possible version of the argument being made by people associated with OpenAI. His purpose is not initially to dismiss their concerns, but to identify exactly where he thinks their reasoning becomes flawed.

The first part of their argument, in his view, is reasonable.

OpenAI and similar frontier laboratories spend enormous amounts of money developing advanced models. These expenses include acquiring computing infrastructure, purchasing or renting access to specialised AI chips, assembling training datasets, paying researchers and engineers, conducting evaluations, and operating the models after they have been trained.

To finance the next generation of models, these companies need to earn sufficiently high returns from the models they have already built. One of OpenAI’s principal business models is therefore to sell access to its models through metered inference: customers pay according to how many tokens they process through the model.

A token is a small unit of text or data handled by a language model. When a customer sends a prompt and receives an answer, both the input and output consume tokens. OpenAI charges the customer more for those tokens than it costs OpenAI to generate them. That difference is the gross margin from inference.

Stokes accepts that OpenAI probably needs healthy inference margins if it wants to continue financing extremely expensive training runs under its present corporate structure. If inference becomes a low-margin commodity, OpenAI may struggle to generate enough cash internally or convince investors that future spending will produce attractive returns.

The second part of the argument is also plausible.

Powerful open-weight models can make it harder for proprietary model providers to maintain high prices. An open-weight model allows other companies to download or host the model themselves, modify it, fine-tune it, and offer competing services without paying the original developer for every query.

Open models can also facilitate distillation. Distillation broadly refers to using the outputs or behaviour of a highly capable model to help train a smaller or cheaper model. If competitors can reproduce a significant portion of a frontier model’s useful capabilities at much lower cost, the original developer may lose some of its pricing power.

Customers may then ask why they should pay a premium to OpenAI when a cheaper open model, Chinese model, locally hosted model, or specialised model can perform the same task adequately. As alternatives improve, proprietary inference prices may fall toward the underlying cost of providing the computation.

Stokes therefore agrees that open models, cheap imitations, and distillation could materially damage OpenAI’s current economics.

Where he strongly disagrees is with the next step in the argument: the suggestion that if OpenAI cannot make enough money selling tokens, then nobody will have an economic reason to finance future frontier-model training.

Stokes regards this as a major conceptual error. It confuses one company’s chosen method of monetisation with the total economic value created by AI.

Selling metered tokens is only one way to make money from software. In the broader technology industry, foundational technologies are often monetised indirectly. Companies may give away or cheaply provide one layer of technology because it increases the value of another product or service they control.

For example, a company might use AI to improve cloud-computing demand, advertising performance, enterprise software, cybersecurity, healthcare administration, customer service, logistics, financial analysis, legal work, or industrial automation. It does not necessarily need to earn a large margin from every individual token. It may earn its return from the larger product or workflow in which the tokens are embedded.

This distinction is central to Stokes’s argument.

Suppose an AI model makes it possible to automate a task that previously required ten employees. The economic value of that automation may be very large. A software company could charge the employer for the completed workflow, for a subscription to the automation platform, for integration services, for performance improvements, or for a share of the resulting cost savings.

The customer does not necessarily care how many tokens were consumed. The customer cares whether the work was completed accurately and economically.

The profitable product may therefore be:

• An AI customer-service system that resolves support tickets;

• An accounting platform that automatically reconciles transactions;

• A legal system that reviews contracts;

• A coding platform that completes software-development tasks;

• A medical-administration system that processes claims;

• A sales platform that identifies and follows up with prospective customers;

• An autonomous agent that performs an entire business process.

In all these cases, inference is an input into the product rather than the final product itself. The model may be analogous to electricity, cloud computing, database storage, or an operating system: an essential technological component whose value is captured elsewhere in the stack.

Stokes’s response to OpenAI is therefore roughly as follows:

If open competition makes the business of selling raw inference tokens unattractive, that may indeed be bad for OpenAI. But it does not follow that AI itself has become economically unattractive. It may simply mean that the profits will move from the model layer to the application, distribution, integration, data, workflow, or infrastructure layers.

The market is not obligated to preserve OpenAI’s preferred business model. If customers value complete workplace automation more than access to a generic model API, then AI companies will have to build and sell automation products rather than rely mainly on high-margin token sales.

Stokes also rejects the idea that the United States government should protect frontier laboratories from open-source or open-weight competition merely because those laboratories need large margins to finance future research.

From his perspective, asking the government to suppress open models would amount to asking public policy to protect a particular set of companies from technological competition. He believes the more appropriate response is for those companies to adapt their products and business models.

His broader point is not that frontier training is inexpensive or that OpenAI’s financial concerns are imaginary. It is that the economic value of advanced AI is much larger than the market for standalone inference tokens. Even if token margins collapse, companies may continue financing model development because the models increase profits elsewhere.

A cloud provider might subsidise model training because better models generate more cloud demand. An enterprise-software company might fund training because it can charge more for AI-enabled productivity tools. A consortium of large companies might jointly finance a model because all of them benefit from having access to it. Governments, universities, defence organisations, or industrial groups might also fund model development for strategic reasons.

Therefore, the collapse of the “token merchant” business model would not necessarily mean the end of frontier-model development. It might simply change who finances the models, how they are financed, and where the resulting profits are captured.

Nic Carter’s argument: The American economy does not depend on OpenAI or Anthropic preserving their margins

Nic Carter agrees with Stokes and extends the argument into questions of industrial policy, geopolitical competition, and beliefs about artificial general intelligence.

In Carter’s terminology:

• “OAI” means OpenAI.

• “Ant” means Anthropic.

• “USG” means the United States government.

• “Token merchants” refers to companies whose business depends heavily on selling metered access to model inference.

• “Neoclouds” refers to newer cloud infrastructure companies focused heavily on supplying GPU computing for AI.

• “Hyperdeflation in the cost of digital cognition” means that the cost of obtaining machine-generated reasoning, analysis, writing, coding, and other intellectual work is falling rapidly.

Carter’s starting point is straightforward: the United States government does not owe OpenAI or Anthropic a permanently profitable business model.

If cheaper Chinese models, open-weight releases, distillation, or intense competition reduce the price of AI inference, this may harm the margins and valuations of the leading American laboratories. But it does not necessarily harm American consumers or businesses.

In fact, cheaper AI could benefit almost every downstream user.

Consumers would obtain better and less expensive AI services. Enterprises could integrate AI into more workflows because the economics would become more attractive. Software companies could build products that would have been uneconomical when inference was expensive. Start-ups would face lower barriers to entry. Cloud and data-centre providers could experience greater demand as falling prices stimulate additional usage.

This is what Carter means by “hyperdeflation in the cost of digital cognition.” Intellectual capabilities that were once scarce and expensive—such as drafting documents, analysing data, writing code, translating languages, or answering technical questions—could become abundant and inexpensive.

From the perspective of the overall economy, this could be highly beneficial even if it destroys some of the economic rents previously earned by frontier laboratories.

Carter therefore asks: why should society be especially worried if the companies selling raw tokens earn lower profits?

The AI infrastructure providers could still sell computing capacity. Internet companies could incorporate cheaper models into their products. Enterprises could continue automating work. Consumers could obtain more intelligence for less money. The total use of AI might increase substantially as its price falls.

The main losers would be OpenAI and Anthropic in their current forms, especially if their valuations, financing plans, and future research budgets depend on maintaining high inference margins. But even those companies would not necessarily disappear. They could adapt by selling enterprise software, specialised agents, security, custom models, data services, integration, or outcome-based automation rather than relying primarily on token markups.

The hidden assumption: Does one laboratory need to “win forever”?

Carter argues that the fear surrounding falling token margins makes sense only if one accepts a much stronger and more speculative belief about the development of AI.

That belief is that the identity of the leading AI laboratory at the moment of a major capability breakthrough will determine the long-term future of the world.

This is the idea of strong path dependence.

Under a strongly path-dependent view, it is critically important which organisation reaches transformative AI first. If one laboratory develops a system capable of recursive self-improvement—often abbreviated as RSI—the system might improve its own research and engineering abilities, which would allow it to design an even better system, which would then improve itself again.

If this feedback loop were sufficiently powerful, the first company to initiate it might achieve an overwhelming and possibly permanent advantage. It could become impossible for competitors to catch up.

Carter mocks this scenario by describing it as a laboratory that “logs on and wins forever.” His reference to an “AI rapture” is intended to compare some AGI narratives to a religious end-times belief: at a decisive moment, one chosen organisation creates superintelligence and history enters an entirely new phase.

His joking phrase “GEOTUS Dario” refers to Anthropic CEO Dario Amodei, portraying him sarcastically as a future supreme ruler following an Anthropic-led intelligence explosion.

Carter is not necessarily claiming that nobody at OpenAI or Anthropic has legitimate safety concerns. His point is that their economic arguments may be influenced by a belief that their continued leadership is uniquely important for humanity.

If Anthropic believes it is more responsible than competing laboratories, then a loss of market leadership is not merely a commercial problem. It becomes a safety problem: a less responsible company or a geopolitical rival might reach transformative AI first.

Similarly, if OpenAI sees itself as the organisation most likely to build beneficial AGI, then anything that weakens OpenAI’s ability to finance the next training run can be framed as a threat to technological progress or national security.

Under that worldview, protecting the leading laboratories’ margins may appear justified because their survival is supposedly necessary to ensure that the “right” organisation wins the AI race.

Carter rejects, or at least heavily discounts, this assumption. He believes advanced AI will transform society, but he is sceptical that progress will take the form of a sudden, winner-takes-all intelligence explosion.

He expects real-world diffusion to be slower and more complicated. Organisations must integrate models into existing systems, change workflows, obtain regulatory approval, protect sensitive data, manage errors, retrain workers, and redesign business processes. Even very capable AI may take years to spread across the economy.

If there is no immediate “winner takes everything” moment, then it matters less whether OpenAI or Anthropic retains the highest market share at every stage of development. Competitive pressure that lowers prices may then be economically healthy rather than existentially dangerous.

Why Carter views Chinese open models as beneficial to AI users

Carter argues that Chinese laboratories are helping commoditise the model layer by releasing capable open-weight models at relatively low prices.

This creates margin pressure for American frontier laboratories. It makes it harder for them to charge large premiums for capabilities that competitors can approximate cheaply.

From the perspective of OpenAI or Anthropic investors, that is bad. From the perspective of AI buyers, however, it is good. Enterprises, developers, governments, and consumers gain more bargaining power. They can choose among more suppliers, host models themselves, negotiate lower prices, and avoid becoming completely dependent on a small number of closed-model providers.

Carter therefore describes China as doing “everyone a favour” by reducing the pricing power of the leading laboratories.

This does not mean he is indifferent to geopolitical competition. He says he would prefer an American organisation rather than a Chinese organisation to achieve a decisive recursive-self-improvement breakthrough. But he also considers a world dominated permanently by one or two American closed-model companies undesirable.

His preferred outcome appears to be a competitive and decentralised AI market in which the benefits of cheap intelligence are broadly distributed, rather than a closed oligopoly in which OpenAI and Anthropic control access and extract very high margins.

Carter’s response to the objection that China is defeating the United States

One obvious objection is that celebrating successful Chinese open models amounts to celebrating an American strategic defeat.

Carter’s answer is that this may be a limited Chinese victory, but it is not necessarily a durable or comprehensive one.

The United States continues to possess major structural advantages, including advanced AI chips, leading semiconductor-design companies, hyperscale cloud platforms, large data-cententre capacity, specialised AI infrastructure providers, capital markets, and many of the strongest frontier models.

Chinese laboratories may weaken the commercial economics of American model developers without building a sustainable business for themselves. Releasing an open-weight model can destroy a competitor’s pricing power, but it does not automatically generate revenue for the organisation that released it.

Carter contrasts this with Chinese industrial strategies in sectors such as solar panels, batteries, and drones. In those industries, Chinese companies used scale, government support, manufacturing efficiency, and aggressive pricing to capture substantial market share and build large domestic industries.

Open-weight AI may be different. A Chinese laboratory might release a model that reduces OpenAI’s prices, yet earn relatively little from the release. The economic value could instead be captured by American cloud providers, semiconductor companies, enterprise-software companies, and businesses using the model.

In that case, China may have damaged the profitability of certain American laboratories without taking control of the broader AI economy.

Carter also suggests that the United States could respond symmetrically. If a Chinese company later tries to monetise a closed proprietary model, American laboratories could release competitive open models and weaken the Chinese company’s pricing power.

Most importantly, the United States could continue benefiting from AI even if open-weight models become dominant. American companies could host those models, optimise them, integrate them into products, and run them on American infrastructure.

The United States might therefore retain an aggregate advantage in computing capacity and AI deployment even without a single American laboratory achieving permanent monopoly power.

Carter’s response to the objection that nobody will finance new models

A second objection is that if distillation and open models eliminate frontier-model profits, companies will stop carrying out expensive training runs. Innovation will stall because every new model will simply be copied.

Carter considers this concern overstated.

First, if distillation truly made further large training runs economically irrational for everyone, the result would not necessarily give China an advantage over the United States. It could create a technological stalemate in which neither side has a strong incentive to spend enormous sums pushing the frontier.

Second, AI development may be shifting away from relying exclusively on ever-larger pretraining runs. Future progress could come from post-training, reinforcement learning, synthetic data, tool use, improved reasoning methods, specialised fine-tuning, more efficient architectures, and models designed for particular categories of work.

The frontier may therefore advance through many smaller or more targeted investments rather than only through enormous general-purpose pretraining runs.

Third, Carter doubts that all frontier training will stop. Developers may find technical or commercial ways to reduce the threat of distillation. They might restrict access to their most capable models, monitor suspicious usage, impose rate limits, permit access only to verified customers, or design systems whose full capabilities are difficult to reproduce from outputs alone.

There could also be new financing structures. Several technology companies, cloud providers, governments, or industrial groups could form a consortium to share the cost of developing a model. Each participant might accept a lower direct financial return because it benefits strategically from access to the resulting technology.

A cloud company might finance model development because the model increases demand for computing. An enterprise-software company might participate because the model improves its products. A government might contribute for defence, research, or national-security reasons.

In other words, model training does not necessarily need to be financed solely by the profit earned from selling the model’s tokens.

Carter’s response to the claim that OpenAI and Anthropic are already essential economic infrastructure

A third objection is that OpenAI and Anthropic have become so central to the AI investment cycle that severe financial trouble at either company would damage the broader economy.

Carter accepts that their collapse or sharp devaluation would hurt certain groups.

Investors could suffer large losses. Planned public offerings might be cancelled. Companies that had built their valuations around contracts with major AI laboratories could be marked down. Some specialised cloud providers could face difficulties if a large part of their contracted future revenue depended on OpenAI or Anthropic.

The reference to “RPOs” concerns remaining performance obligations: contracted revenue that a company expects to recognise in the future. If an AI infrastructure provider has large long-term contracts with a laboratory, investors may treat those contracts as evidence of future growth. If the laboratory becomes unable to honour them, the infrastructure provider could suffer.

Nevertheless, Carter does not believe OpenAI or Anthropic is indispensable to the American economy.

Microsoft, Google, Amazon, Meta, and other large technology companies have their own AI capabilities, cloud platforms, research teams, data, products, and access to capital. They could continue investing in AI even if one or both independent frontier laboratories were severely weakened.

Most ordinary households also have little direct financial exposure to OpenAI or Anthropic. Losses would be concentrated among particular investors and commercial partners rather than distributed throughout the population in the manner of a systemically important bank failure.

Moreover, if AI inference became dramatically cheaper, demand for AI computing might rise rather than fall. This is an application of the idea commonly associated with Jevons paradox: when a technology becomes more efficient and less expensive, total consumption can increase because many more uses become economical.

Thus, even if laboratories earned less revenue per token, the world might consume vastly more tokens. Data centres, cloud providers, application developers, and enterprises could continue expanding AI deployment.

Carter’s conclusion is that the AI investment boom is not necessarily dependent on preserving the profit margins of OpenAI and Anthropic. The capital spending will still create computing capacity, and that capacity can serve many models and customers.

The central message shared by Stokes and Carter

Both writers are making a distinction between three different things that are sometimes treated as though they were identical:

1. The profitability of selling access to proprietary AI models;

2. The economic value generated by AI throughout the economy;

3. The continued advancement of AI capabilities.

OpenAI and Anthropic may need high token margins to sustain their current valuations and finance future training under their existing business models.

But Stokes and Carter argue that it does not follow that society needs OpenAI and Anthropic to maintain those margins.

Cheap and abundant models could reduce the profits captured by the model developers while increasing the overall economic value created by AI. The benefits would move downstream to software companies, cloud providers, enterprises, and consumers.

Frontier research might then be financed differently: through cloud revenues, application profits, enterprise subscriptions, strategic investment, government funding, industry consortia, or vertically integrated technology companies.

The fundamental disagreement is therefore not about whether frontier training costs money. Everyone accepts that it does.

The disagreement concerns who must pay for it, where the profits must be earned, and whether the survival of today’s leading laboratories in their present forms is essential to continued AI progress.

Stokes and Carter believe it is a mistake to treat OpenAI’s or Anthropic’s current business model as a public good that must be protected from competition. If selling raw inference tokens becomes unprofitable, those companies should adapt just as previous technology companies have had to adapt when an important product became commoditised.

Their broader vision is that AI becomes an inexpensive, widely available input into the economy. In this world, the model layer may have lower margins, but the applications built on top of it become far more numerous and valuable.

Where their argument remains uncertain

Their argument is coherent, but it depends on several unresolved questions.

It is uncertain whether downstream applications will generate enough profit to finance the most expensive frontier research. A company can benefit from cheap models without wanting to contribute billions of dollars toward building the next generation.

There may therefore be a collective-action problem: everyone wants to use advanced models, but each participant would prefer someone else to bear the cost of developing them.

It is also uncertain how easily frontier models can be distilled. If inexpensive models can reliably copy most valuable capabilities, the incentive to conduct the original training could weaken substantially. Conversely, if frontier capabilities are difficult to reproduce, leading laboratories may retain more pricing power than Stokes and Carter assume.

Another uncertainty concerns timing. Even if a new ecosystem of consortia, vertically integrated applications, or government-supported research eventually emerges, the transition could be disruptive. A sudden collapse in laboratory financing might temporarily slow model development and damage infrastructure companies whose expansion plans were based on contracts with those laboratories.

Finally, the importance of market leadership depends on the technical path AI follows. If progress is gradual and widely distributed, Carter’s competitive-market argument looks stronger. If recursive self-improvement creates a sudden and overwhelming first-mover advantage, then the identity and governance of the leading laboratory could matter enormously.

That technical question remains unsettled.

Overall interpretation

Jon Stokes is primarily making a software-industry and business-model argument:

The inability to earn high margins from raw inference does not mean AI lacks economic value. It means the model provider may be trying to capture value at the wrong layer.

Nic Carter expands this into an industrial-policy and ideological argument:

The United States should not protect OpenAI or Anthropic merely because their preferred token-selling model is threatened. Cheaper intelligence may be good for the broader economy, even if it is bad for the laboratories’ investors.

Carter further argues that much of the alarm about weakening the leading laboratories arises from a belief that one of those laboratories must reach transformative AI first. People who believe in a sudden winner-takes-all intelligence explosion naturally treat OpenAI’s or Anthropic’s continued leadership as an overriding national and civilisational priority.

People who expect AI to develop and diffuse more gradually are more likely to welcome competition, falling prices, open models, and reduced concentration of power.

The debate is therefore about much more than API prices. It concerns whether AI will resemble a competitive technology industry in which value shifts among different layers, or a race toward a singular breakthrough in which the first laboratory to reach the finish line acquires a permanent and potentially world-changing advantage.

On Friday, I published a video discussing the latest market decline. To be precise, the Nasdaq Composite fell about 1.4% during the Friday, 17 July U.S. trading session.

Kimi K3 was probably one catalyst for the sell-off, especially among semiconductor and AI-related stocks. However, it was not the only factor. Investors were already concerned about stretched AI valuations, the returns on enormous data-centre investments, corporate earnings and geopolitical risks. It is therefore too simplistic to attribute the entire market decline to one new AI model.

During my livestream, a viewer asked whether I had watched Mr. Loo’s discussion of Kimi K3. I had not watched it at that point, so I could not offer a detailed response.

I have since watched the relevant portion of his livestream, published on 18 July. There are several claims that I believe require clarification.

Let me start by saying that Kimi K3 is genuinely impressive. Moonshot AI deserves credit for building a model that approaches the capabilities of the best proprietary American models.

But recognising Kimi K3 as an important achievement does not mean we should jump directly to the conclusion that American AI companies, or every stock associated with AI, will collapse.

1. Open-weight does not automatically mean extremely cheap

Mr. Loo described Kimi K3 as open source, meaning that users can download its parameters and customise the model. He then connected this directly to Kimi’s lower price.

There are a few different concepts being mixed together here.

First, Kimi K3 is more accurately described as an open-weight model, rather than fully open-source software. Open-weight means that the trained model parameters are made available. It does not necessarily mean that the complete training data, training code, data-cleaning process and every part of the development pipeline are openly disclosed.

There is also a timing issue. As of 19 July 2026, Kimi K3’s weights are not yet available for download. Moonshot says that the full weights will be released by 27 July, together with additional technical information. Artificial Analysis therefore still classifies the model as proprietary for the moment.

Second, open weights do not make inference free.

With a closed model, users generally access the service through a subscription or an API. The provider pays for training, servers, networking, electricity, engineering staff and continuing model development. It then tries to recover those costs through subscriptions or usage charges.

With an open-weight model, a company may eventually download the weights and operate the model on its own infrastructure. This may eliminate the need to pay the model developer for every token, but the cost does not disappear. It is transferred to the user.

The user must now pay for:

• AI accelerators;

• servers and high-speed networking;

• electricity and cooling;

• inference software;

• engineers to deploy and maintain the system; and

• spare capacity to handle peak demand.

This is especially important for Kimi K3 because it has 2.8 trillion total parameters. Moonshot itself recommends configurations containing at least 64 accelerators connected within a high-bandwidth domain. This is not a model that an ordinary consumer can download and run comfortably on a personal computer.

Open weights can still lower costs by increasing competition, allowing customisation and reducing dependence on a single API provider. But we cannot assume that Kimi’s API price is low simply because the model will eventually be open-weight. Its price may also reflect architectural efficiency, serving optimisation, competitive strategy or even deliberate subsidisation. Moonshot has not provided enough information for us to separate these factors.

Price per token is not cost per task

There is another major problem with comparing AI costs: price per million tokens is not the same thing as the cost of completing a task.

Different models may consume dramatically different numbers of tokens to solve the same problem. A model charging half as much per output token could still cost almost the same per task if it generates twice as many tokens.

Artificial Analysis offers a better like-for-like comparison. Its Intelligence Index calculates the average cost of completing its evaluation tasks after considering input tokens, cached input, reasoning tokens and answer tokens, not simply the advertised output price.

Here is the comparison between Kimi K3 and GPT-5.6 Sol at maximum reasoning:

Kimi’s advertised output-token price is half that of GPT-5.6 Sol. However, Kimi generated approximately 1.86 times as many output tokens during the Intelligence Index evaluation.

After taking actual token consumption into account, Kimi’s average cost per task was only around 9.6% lower—US$0.94 versus US$1.04. GPT-5.6 Sol also achieved the slightly higher intelligence score of 59 versus Kimi’s 57.

That is still a respectable result for Kimi. But it does not support the claim that Kimi provides the same overall capability at half, one-third or one-quarter of GPT-5.6 Sol’s cost.

It is possible to obtain very different comparisons by selecting particular benchmarks or workloads. Kimi is especially strong in certain coding and agentic tasks. But for a broad statement about general model capability and cost, cost per completed task is far more meaningful than price per token.

2. There is no evidence that Kimi K3 uses Huawei chips instead of Nvidia chips

Mr. Loo also stated that Kimi is cheaper because it does not use Nvidia chips and apparently uses Huawei chips.

At present, this claim is unsupported.

Moonshot has not disclosed that Kimi K3 was trained on Huawei Ascend processors. It has not disclosed that the model is primarily served using Huawei hardware either.

The Kimi launch article mentions Nvidia H200 GPUs and a “GPGPU from an alternative vendor,” but this appeared in a test of Kimi K3’s ability to optimise GPU kernels. It was not a disclosure of the hardware used to train the model.

The unnamed alternative vendor could potentially be Huawei, but it could also be another supplier. Until Moonshot identifies it, presenting Huawei as a fact is speculation.

Moonshot says that more information about K3’s architecture and training will be included in the forthcoming technical report. Until that report is published, the most accurate conclusion is:

We do not currently know the complete hardware configuration used to train Kimi K3, and there is no public evidence that it was trained primarily on Huawei chips.

There is an additional contradiction in the argument. If Kimi K3 had achieved frontier performance primarily on Huawei chips, that would not make the model “nothing serious.” It would be a major technological development for both China’s AI industry and Huawei.

But we should wait for evidence before making that claim.

3. DeepSeek was not invalidated by distillation

Mr. Loo compared Kimi K3 with the DeepSeek episode of January 2025. He suggested that investors initially panicked, but later discovered that DeepSeek had used distillation and concluded that DeepSeek was “nothing big deal.”

This mixes together several different concepts.

DeepSeek-R1 was released on 20 January 2025. On 27 January, Nvidia’s share price fell by approximately 17%, wiping close to US$600 billion from its market capitalisation. At that time, it was the largest single-day destruction of market value experienced by any listed company.

The market reaction was unquestionably extreme. But the later recovery in Nvidia’s share price does not prove that DeepSeek was technologically insignificant.

What did DeepSeek actually distil?

DeepSeek released its main R1 and R1-Zero reasoning models together with six smaller models distilled from R1.

The smaller models were based on Qwen and Llama architectures and were trained to absorb some of the reasoning behaviour demonstrated by the larger R1 model. But the flagship DeepSeek-R1 model itself was not simply a small model copied from another company through distillation.

According to DeepSeek’s paper, R1 was developed using cold-start data, supervised training and multiple stages of reinforcement learning. DeepSeek-R1-Zero demonstrated that substantial reasoning behaviour could emerge through large-scale reinforcement learning even without conventional supervised fine-tuning at the beginning.

Knowledge distillation is also not some embarrassing secret that automatically discredits a model. It has been a standard machine-learning technique for more than a decade. Google researchers formally described modern neural-network distillation in 2015, and OpenAI has openly offered model-distillation tools to developers.

The relevant questions are therefore not merely whether distillation was used, but:

• What was distilled?

• From which teacher model?

• How much independent training was performed?

• What capabilities were retained?

• How efficiently can the resulting model operate?

Treating the word “distillation” as evidence that DeepSeek was somehow fake misses the technical substance.

DeepSeek’s innovations remain important

DeepSeek’s contribution was not limited to releasing an inexpensive chatbot.

DeepSeek-V3 used a sparse mixture-of-experts architecture containing 671 billion total parameters but activating only about 37 billion for each token. Instead of running the entire model for every token, the system routes each token through a smaller selection of specialised experts.

This reduces the amount of computation required relative to activating every parameter.

DeepSeek also used Multi-head Latent Attention, or MLA. MLA compresses information stored in the key-value cache, reducing memory requirements during inference. This is especially valuable when processing long conversations or serving large numbers of users simultaneously.

These were meaningful architectural and engineering advances. The fact that Nvidia subsequently recovered does not erase them.

The more plausible explanation is that model efficiency and demand for computing power are not opposites.

Model developers used efficiency improvements to build more capable models, support longer contexts, perform more reasoning and operate increasingly autonomous agents. The saved computation was not necessarily returned to investors as permanently lower infrastructure spending. Much of it was reinvested into doing more ambitious work.

4. Kimi K3 does not automatically mean AI stocks will collapse

Mr. Loo’s strongest conclusion was that if Kimi K3 is genuinely good, AI stocks will “collapse for sure” and the AI bubble may burst.

This conclusion does not follow from the evidence.

Kimi K3 can be simultaneously:

1. a major technological achievement;

2. a competitive threat to some model providers;

3. a source of pricing pressure on AI APIs; and

4. positive for the long-term adoption of AI.

The mistake is to assume that cheaper intelligence must result in less demand for AI hardware.

The missing concept: Jevons Paradox

Jevons Paradox describes a situation in which improvements in efficiency reduce the cost of using a resource, causing demand to expand so much that total consumption increases rather than decreases.

The principle does not guarantee that every efficiency improvement will increase total consumption. Demand must be sufficiently responsive to lower prices and better capabilities. But AI appears to have many of the conditions under which a strong rebound effect could occur. Recent economic research has specifically examined how falling inference prices could encourage companies to adopt more compute-intensive AI agents and workflows.

Suppose a company can reduce the cost of an AI task from ten dollars to one dollar.

It may initially save nine dollars. But it may then decide to run the task 100 times instead of once. It may also introduce the model into departments and applications where AI was previously too expensive or not capable enough.

Cheaper and better models can unlock:

• autonomous software agents working for hours rather than answering one question;

• AI-generated video and interactive media;

• continuous monitoring of business operations;

• personalised education for millions of students;

• large-scale scientific simulations;

• automated software testing;

• real-time translation; and

• AI assistants for employees and consumers.

Many of these applications would be uneconomical if each interaction were expensive. Improving efficiency does not merely reduce the cost of existing usage. It creates entirely new categories of demand.

Kimi K3 itself illustrates this dynamic. It activates only 16 of its 896 experts for each token and incorporates new attention mechanisms intended to improve scaling efficiency. But Moonshot did not use those gains to build a tiny model requiring almost no infrastructure. It used them to construct a 2.8-trillion-parameter model with a one-million-token context window, native visual capabilities and support for long-running agentic work.

In other words, the efficiency improvement was converted into greater capability.

That is Jevons Paradox in action, or at least the mechanism through which it could occur.

Open models accelerate the entire AI industry

Open-weight models also allow the wider research community to study, adapt and improve upon new techniques.

Closed-model companies such as OpenAI and Anthropic generally disclose only part of their training and optimisation methods. Their most valuable techniques remain proprietary.

When an open-weight developer publishes its model, architecture and technical findings, researchers around the world can study what worked. Competitors may incorporate similar ideas into future models, while inference providers can optimise their software and hardware around the architecture.

This accelerates the rate at which intelligence becomes cheaper and more capable.

That does create competitive pressure. Closed-model providers cannot assume that high margins will last forever. They must continue improving their models, developer tools, reliability, safety, enterprise integration and user experience.

However, greater competition among model providers does not necessarily reduce the total AI market. It may expand it.

The internet did not collapse because the cost of transmitting information fell. Cloud computing did not disappear because servers became more efficient. Smartphones did not reduce semiconductor demand by making computing more accessible.

Lower cost frequently expands the market.

“AI stocks” are not one single investment

It is also unhelpful to discuss all AI stocks as though they have identical business models.

Kimi K3 may affect different parts of the AI ecosystem differently.

Proprietary model providers could face stronger pricing pressure. Smaller AI laboratories may find it harder to compete with frontier open-weight models. Cloud providers could benefit from additional inference workloads. Software companies could build more capable applications at lower cost. Semiconductor companies could experience greater total demand if AI usage expands faster than efficiency improves.

Even within semiconductors, the impact will differ across GPUs, networking equipment, memory, storage, power systems and custom accelerators.

Therefore, a new model cannot logically mean that every AI-related company will “collapse for sure.”

Investors still need to examine each company’s valuation, competitive advantage, margins, capital requirements and ability to convert AI spending into profits. Jevons Paradox does not mean that every AI stock is fairly valued, nor does it guarantee that every company will win.

It means only that better efficiency is not automatically bearish for total computing demand.

My conclusion

I agree with Mr. Loo on one important point: Kimi K3 should be taken seriously.

It shows that Chinese AI laboratories are operating close to the global frontier. It demonstrates that open-weight models can compete with expensive proprietary systems on sophisticated coding and knowledge-work tasks. It could put pressure on API prices and force American AI companies to innovate faster.

But several conclusions in his discussion are not supported by the available evidence.

Kimi K3 is not currently available for users to download, although Moonshot says its weights will be released by 27 July. Its open-weight strategy is not, by itself, the reason its API is cheaper. There is no evidence that the model was trained primarily using Huawei chips. DeepSeek was not invalidated because it released distilled models. And an improvement in model efficiency does not automatically imply the collapse of AI infrastructure demand.

The most important question is not:

“How much cheaper is each token?”

It is:

“How many additional tasks will people perform once intelligence becomes cheaper and more capable?”

If AI prices fall while the number and complexity of AI tasks grow even faster, total demand for computing power can continue rising.

Kimi K3 may indeed be another DeepSeek moment.

But the lesson from DeepSeek should not be that efficient models are insignificant. The lesson is that the AI industry can improve its algorithms rapidly, and then use those improvements to pursue even larger models, longer reasoning, more autonomous agents and many more applications.

That is not necessarily the end of the AI investment cycle.

It may be one of the forces extending it.

Information and benchmark results are current as of 19 July 2026. Moonshot has said that Kimi K3’s full weights and further technical information will be released by 27 July, so some conclusions may need to be revisited after those disclosures.

Earlier today, a Telegram group member shared these three pages from Dow 36,000 in my Telegram group. The excerpts contain extraordinarily bullish projections for Cisco, Microsoft and Biogen near the height of the dot-com boom.

Read with more than two decades of hindsight, some of the assumptions now look almost comical.

The screenshots were shared in the context of today’s excitement around artificial intelligence, with the implied message being straightforward: investors were overly optimistic about technology in 1999, and they are making the same mistake again with AI.

It is an entertaining comparison, and historical mistakes are always easy to recognise after the outcome is known. But I think the more interesting question is not whether those old forecasts were wrong. They clearly were. The question is whether finding a failed bullish forecast from the past actually tells us much about the opportunities and valuations in front of us today.

Here is my view.

A historical analogy is not an argument

It is always entertaining to reopen an old book, find a spectacularly optimistic forecast that turned out to be wrong, and make fun of it with the benefit of 25 years of hindsight.

But we should be clear about what this exercise actually proves.

Posting pages from Dow 36,000 alongside today’s AI valuations is reasoning by analogy. You are identifying a historical episode that resembles the present and using that resemblance to support your conclusion: people were excessively optimistic about technology in 1999, so people who are optimistic about AI today must be making the same mistake.

There is nothing inherently wrong with reasoning by analogy. We all do it. Thinking from first principles every time would be exhausting. In one limited sense, the human mind works somewhat like a language model: it absorbs past examples, recognises patterns and uses those patterns to interpret new situations.

The danger is that pattern recognition is highly vulnerable to selection bias. Once you have concluded that AI is another dot-com bubble, you will naturally search for historical examples that confirm it. You will find Cisco forecasts, failed internet companies and absurd valuation models. You are much less likely to search for pessimistic forecasts about technology companies that subsequently became enormously successful.

But those examples are equally easy to find.

Consider Facebook’s IPO in 2012. The company was valued aggressively relative to its then-current financial results. It had generated only $3.7 billion in revenue and $1 billion in net income in 2011, while advertising accounted for 84% of revenue. (WIRED)

A contemporary Wired article captured several prominent bearish arguments. Paul Kedrosky argued that much of Facebook’s value had already been captured in private markets and expected its post-IPO performance to resemble disappointing listings such as Groupon and Zynga. Forrester analyst Nate Elliott questioned whether Facebook advertising produced adequate returns for marketers. He characterised its advertising model as simplistic and compared Facebook with Yahoo rather than Google. The article also raised the possibility that privacy regulation, user resistance and weak advertising effectiveness could limit Facebook’s monetisation. (WIRED)

Those concerns were not irrational. They were based on the information available at the time. But the conclusions proved far too pessimistic.

By 2025, Meta’s annual revenue had reached approximately $201 billion, more than 50 times Facebook’s 2011 revenue, while net income was about $60.5 billion. (Meta) In the first quarter of 2026 alone, Meta generated $55 billion of advertising revenue. (SEC) The supposedly simplistic advertising business became one of the most powerful monetisation engines ever created.

Of course, this Facebook example does not prove that today’s AI companies are fairly valued. It does not prove that AI spending will earn adequate returns, or that every AI-related company will succeed. Using Facebook to declare that there is no AI bubble would simply replace one historical analogy with another.

That is precisely the point.

For every old bullish forecast that looks foolish today, someone can usually find an old bearish forecast that looks equally foolish. Selecting only one category tells us more about the person’s current belief than about the future.

The proper question is therefore not, “Does this remind me of 1999?” It is:

What are the actual economics of the companies today?

How large is the addressable market? Are customers receiving measurable value? Is demand recurring or experimental? How quickly are inference and infrastructure costs falling? What are the competitive advantages? How much future growth is already reflected in the valuation? And what assumptions must be true for the current price to generate an acceptable return?

Those questions require considerably more work than posting an embarrassing page from an old book. But they also provide a mechanism for changing your mind when the evidence changes.

Historical comparisons can be useful starting points. They are not conclusions. The real danger begins when we collect only the analogies that reinforce what we already believe. Without a corrective mechanism, the same belief can be confirmed repeatedly, even when the underlying reasoning is wrong.

The “we already have AGI” argument — Ghodsi’s full reasoning

Ghodsi’s claim has two parts that are easy to confuse. The first is a definitional point about AGI itself. The second is a diagnostic point about why AI doesn’t feel transformative inside companies even if his AGI claim is true.

Part 1: The definitional move

Ghodsi grounds his AGI claim in the 2009-era research definition, not today’s pop-culture one. He says he was at Berkeley in 2009 (he names AMPLab; the lab by that name was actually launched in 2011 — its predecessor RAD Lab is what existed in 2009), where Michael I. Jordan and other leading AI researchers worked. Back then, the working definition of AGI was something like: artificial intelligence that is general — i.e. not narrow, not single-task — and human-comparable on broad reasoning.

His test on the audience makes the move concrete:

1. “How many of you think we have AGI already?” → ~10% raise hands.

2. “How many of you think a lot of people you interact with are not as smart as the smartest models you use?” → most hands go up.

3. “Now: how many of you think we don’t have AGI yet?” → silence.

His punchline: by the original definition, frontier models are already smarter than many humans you interact with on most general tasks. That is general intelligence. It is artificial. Therefore, AGI. The reason people deny it, he argues, is that the goalposts have moved — every time a model clears a bar, the bar gets redrawn (e.g. “but it can’t count the Rs in strawberry, therefore not AGI”). What used to be called AGI is now considered “just a chatbot.”

So when he says “we already have AGI,” he’s making a deliberately deflationary point: the AGI label has been rebranded to mean superintelligence (Kurzweil-style: recursive self-improvement, 10% GDP jump, 20% unemployment, UBI for all), and that distortion drives the trillion-dollar arms race for GPUs and data centers, which Ghodsi thinks is unwarranted.

Part 2: The diagnostic — why AGI doesn’t feel like AGI inside companies

This is the more important and more original part of his argument. It connects to the MIT NANDA finding that 95% of enterprise GenAI pilots fail. Ghodsi’s claim: the failure is not because the models are dumb; it is because the models are blind. They lack the context that humans inside organizations have in their heads.

“John or Jane” — the context-holder

His clearest illustration is the long-tenured employee. In every department of every company, there is one person — Ghodsi calls them John or Jane — who has been there 15, 20, sometimes 30 years. When anyone gets stuck, the answer is “go ask John.” If that one person leaves, the department falls apart.

What makes John irreplaceable isn’t formal expertise — it’s an enormous, undocumented body of context:

• Implicit process knowledge: how things actually flow, versus how the org chart says they flow.

• Historical decisions: why a system was built this way (often: a workaround for a 2014 bug nobody remembers).

• Edge-case handling: what to do when a customer in this specific situation calls.

• Social map: who has authority over what, who to escalate to, who to never bother.

• Tacit knowledge: the “feel” for when something is about to break.

None of this is in any document. None of it is in the LLM’s training data. It lives in John’s head, accumulated over a career, and most of it has never been articulated even once.

When you drop a frontier LLM into that company and ask it to “do John’s job,” it can’t. Not because it isn’t smart enough — it would be smart enough if it had the context — but because it’s missing the dataset that matters. The agent makes plausible-looking, confident, wrong decisions, because it’s reasoning over a generic world model instead of this specific company’s world model.

The Databricks support example

Ghodsi makes this concrete with his own product. Databricks support cases come from highly trained data scientists at Databricks customers — smart people with PhDs and big salaries — who got stuck on something hard, like a machine-learning model not hitting the F1 score they need. They’ve already exhausted the obvious solutions. That’s why they’re calling support.

He says they tried every off-the-shelf AI support automation product. All of them failed. As soon as the vendors realized which company they were dealing with, they pulled out: “Whoa, we can’t help you, this is too hard.” Not because the models can’t reason — they can — but because the resolution requires knowing how Databricks’ platform behaves under unusual configurations, what the support engineers have seen across 20,000 customers, what the known-issue list looks like internally, what the playbook is for each failure mode. None of that lives in the LLM’s training data. The Databricks support engineer has it; the AI doesn’t.

His conclusion in that frame: “If the AI could have all the context of how our support engineers at Databricks operate, then the AI could do it. It just doesn’t have it.”

The connector anecdote — same lesson, applied

The Databricks data-connector story is the same point in a different domain. Ghodsi could write a Salesforce connector in two days using AI. The team takes nine months. When he asked why, the team came back: AI shaves at most 1.5 months off, getting them to 7.5. Why? Because the months that matter aren’t the coding — they’re the context-gathering: a full quarter sending PMs to customers to understand how Salesforce is actually used in production, then more time setting up real Salesforce instances to test against.

The breakthrough came from a different engineer who realized you don’t need AI to be smarter — you need to rewire the process so that context flows differently: write requirements in a week and iterate (since rewriting is now cheap), outsource the test-environment setup to specialist firms running in parallel, and parallelize across seven engineers. Result: 7 connectors per quarter instead of 1 in 9 months. AI didn’t do it. Process redesign did. The bottleneck was always context capture and context transfer, not raw intelligence.

The synthesis — the dynamo analogy

Ghodsi ties this to Paul David’s 1990 paper on electrification. When the electric motor arrived around 1880, factories already had centralized steam engines spinning overhead line shafts, with belts running down to each machine. The first thing factories did with the new technology: rip out the steam engine and bolt in an electric dynamo where the steam engine used to be. They captured almost no productivity gains. It took 40 years — until roughly 1920 — for factories to realize they needed to reorganize the entire factory floor around the new technology: distributed unit drive (one motor per machine), single-story horizontal layouts, factories moved out of dense cities. Then productivity surged.

Same with PCs in the Solow-paradox era: Solow famously quipped in 1987 that you could see the computer age everywhere except in the productivity statistics. Why? Because companies were using PCs as glorified typewriters — type a memo, print it, file it, have an assistant index it. They had not rewired the work around what computers actually enable.

Ghodsi’s claim: AI today is in the dynamo phase. Companies are taking the new “engine” (the LLM) and bolting it into the same old workflow — same support tickets, same product-management process, same approval chains, same documentation conventions — and then wondering why nothing changes. The intelligence is there. The wiring is not. “I have the brain, but I have to rebuild the human body — the hands, the legs” is how he puts it. The work that captures the value isn’t building smarter models; it’s the slow, unglamorous work of getting all the John-and-Jane context out of human heads and into the agents, and redesigning processes so the agents can act on it. He thinks this transformation will take many years and looks nothing like what Twitter is currently chasing.

Why this argument matters

The implication, which Ghodsi states bluntly to the room: if his diagnosis is right, then the AI value capture is not in chips, infrastructure, or even foundation models — those are the dynamo. The value is at the application layer, where the painful, company-specific work of context extraction and process redesign actually gets done. That’s why he tells the students to invest the metaphorical $100 at the top of the stack (applications), and why he predicts the trillion-dollar AI companies will look as ridiculous in 2026 as Airbnb (renting your bedroom) and Amazon (selling books online) looked in 2000.

It’s also why he tells the audience to chill out. If the bottleneck is human-organizational rewiring, AGI doesn’t arrive on a Tuesday. It diffuses over years.

At first glance, the US brokerage system can feel less safe than Singapore’s CDP model. In Singapore, your shares are held directly in your own name. There is no intermediary layer, and ownership feels clear and tangible. By contrast, when you invest in US markets, your shares are not registered under your personal name. They are held through a chain of institutions, which creates the impression that you are exposed if something goes wrong.

That impression comes from how the system looks on the surface, not how it actually functions underneath.

Direct Ownership vs Beneficial Ownership

In Singapore’s CDP system, you are the registered owner of your shares. Your name sits directly on the company’s shareholder register.

In the US, the structure is different. When you buy shares through a broker like Interactive Brokers, you are recorded as the beneficial owner, not the registered owner. The shares are held in a central system managed by the Depository Trust & Clearing Corporation, with a nominee entity appearing as the legal owner on company records.

This sounds like a loss of ownership, but it is not. You still retain full economic rights to the shares, including dividends, voting rights, and the ability to sell at any time. The difference lies in how ownership is recorded, not whether you actually own the asset.

Why the System Is Structured This Way

The layered custody model exists for a practical reason: scale.

Modern financial markets process millions of trades daily. If every trade required companies to directly update their shareholder registers, the system would be slow, expensive, and operationally fragile. Instead, the US uses a centralised infrastructure that allows brokers and custodians to track ownership efficiently without constantly rewriting company records.

This structure enables fast settlement, deep liquidity, and global participation. It is not a compromise on safety, but an engineering solution to make markets function at scale.

The Core Protection: Segregation of Assets

The most important safeguard in the US system is the legal separation between client assets and the broker’s own assets.

Your shares are not part of the broker’s balance sheet. They are held on your behalf and must be segregated under regulatory rules. If the broker fails, its creditors cannot claim your securities.

This point is often misunderstood. The broker is a custodian and intermediary, not the owner of your assets.

The Backstop: Investor Protection

On top of segregation, there is an additional layer of protection through the Securities Investor Protection Corporation.

If a broker collapses and there is a shortfall in client assets, SIPC steps in to facilitate the return of securities or compensate investors up to specified limits. It does not protect against market losses, but it exists specifically to handle broker failure scenarios.

This provides a second line of defence in the rare event that segregation alone is insufficient.

What Happens If a Broker Fails

In practice, broker failures are handled through established processes.

In a straightforward case, where records are clean and assets are properly segregated, client accounts are transferred to another broker. Investors may experience some disruption, but their holdings remain intact.

In more complex situations, where records need reconciliation, the process can take longer. Access may be temporarily restricted while assets are sorted and transferred.

In rare cases involving fraud or misuse of assets, there may be a shortfall. This is where SIPC becomes relevant, stepping in to recover or compensate missing assets within its limits.

The key point is that failure does not mean your assets disappear. There is a structured resolution process designed to return them.

Why Institutions Trust This System

Perhaps the strongest evidence of the system’s robustness is who uses it.

The same structure underpins global investing. Pension funds, sovereign wealth funds, insurance companies, and hedge funds all operate within this framework. These institutions manage large pools of capital and apply strict due diligence to counterparty risk.

Their continued reliance on this system reflects confidence in its legal and operational safeguards.

If brokerage risk were as severe as often claimed, global capital markets would not function at their current scale.

Where the Real Risks Actually Are

This does not mean the system is risk-free.

Risks still exist in areas such as broker quality, regulatory standards, and operational controls. Smaller or poorly regulated brokers can pose higher risks. Accounts involving margin or securities lending can introduce additional complexity. In extreme cases, delays or partial losses are possible.

However, these are edge cases rather than the expected outcome, especially when dealing with large, well-regulated brokers.

The Bottom Line

The US system looks more complex than CDP because it relies on intermediaries and layered custody. But complexity does not mean fragility.

Instead of direct registration, the system relies on legal segregation, centralised infrastructure, and investor protection mechanisms to safeguard ownership. This allows markets to operate efficiently at global scale while still protecting investors.

In that sense, the difference is not about safety versus risk. It is about two different ways of achieving the same goal, one through simplicity, and the other through structured layers designed to handle the realities of modern financial markets.

If you’re serious about investing globally, choosing the right broker matters.

Interactive Brokers is one of the most widely used platforms by both retail investors and institutions, with strong regulation, low costs, and access to global markets. If you’re planning to open an account, you can use my referral link below.

https://ibkr.com/referral/boontee954

If you prefer a more intuitive interface and built-in tools for research and trading, Moomoo is a solid option. It’s popular among retail investors for its clean app design and market data. If you’re opening an account, you can use my referral link below.

http://moomoo.com/en-sg/act/welcome?channel=1002&subchannel=268

Currently, my portfolio is roughly 20% ETFs and 80% individual holdings.

I invest in individual stocks not because I am confident I can beat the market index, such as the S&P 500. Rather, I prefer to invest in quality businesses that I believe I understand. Often, I choose companies that I think have less downside rather than those with the most upside.

Another benefit of investing in individual stocks is that I can understand the businesses I invest in at a much deeper level. This is also my way of managing emotions. I am not investing in a black box whose contents I do not understand. These are companies whose business models, products and services, culture, customers, and suppliers I know. Instead of focusing only on price movements, I can monitor many other KPIs, which often reduces my anxiety during market crises.

Mr. Loo mentioned that he invests in a broad-based S&P 500 index, whereas I invest in individual stocks, so the comparison is not apples to apples. Let me clarify:

1. Yes, my investments are mainly in individual stocks, but I do not think one approach is superior to the other. Both have pros and cons. I invest in individual stocks not because I have any unique “kung fu,” but because I prefer owning businesses I can understand. I am willing to accept the risk of underperforming the index. However, since my top holdings are similar to those in major ETFs, I expect that even if I underperform, the gap should not be large.

2. In the screenshot of my performance, I showed a return of 17.14%, which is higher than SPY’s 14.91%. In other periods, my portfolio has underperformed. My point is not that I am particularly skilled at stock picking. Rather, even a simple dollar-cost averaging strategy into SPY could achieve 14.91%, which is a respectable result. For reference, if I had invested a lump sum in the S&P 500 in September 2019 and held until December 2025, the return would have been about 14.2%, slightly lower but not by much.

   

3. I acknowledge that Mr. Loo’s crash-buying strategy produced higher returns, but this depends heavily on the chosen start date. For example, if the start date were October 2023, about 75% of the capital would still be in cash today because only 25% would have been deployed during October 2023 and April 2025. The portfolio return would be significantly dragged down by the large cash allocation. In that case, a lump-sum approach could have produced a 25% return versus roughly 9% for the crash-buying strategy. Rather than dwelling on hypothetical scenarios, I will leave it at that.

Mr. Loo also mentioned that stock investing requires strong emotional management. I understand this point, but the nature of individual stock investing matters. Many people trade frequently to exploit the higher volatility of individual stocks. My approach is different. I identify strong businesses and hold them for the long term. I rarely trade in and out, though I may add to positions gradually through dollar-cost averaging. I do not use technical analysis, and my strategy is primarily buy-and-hold. Over time, holding a basket of individual stocks provides diversification at the portfolio level, so the overall results are not dramatically different from investing in an index. In the long run, I also plan to increase my ETF allocation and reduce individual stock exposure, mainly for estate tax considerations.

Active vs. Passive in Stock Selection and Timing

To help clarify the differences between Mr. Loo’s approach and mine, I developed the diagram below. There are two dimensions of active versus passive investing: stock selection and timing. An investor can be active or passive along either dimension.

• Mr. Loo is passive in selection, since he invests in ETFs, but much more active in timing, deploying capital mainly during market crashes.

• My approach is the opposite. I am active in stock selection but largely passive in timing. I believe long-term returns come primarily from staying invested, not from timing entry and exit points.

• Most traders are active in both selection and timing. In my view, this is the approach to avoid.

• Most dollar-cost averaging investors are passive in both dimensions. I believe about 90% of investors should adopt this approach.

Outperforming the market through stock selection is difficult, and outperforming through timing is equally challenging. You can design sophisticated crash-buying strategies, but even if they produce strong results, it is hard to determine whether the outcome reflects skill or luck. If you judge success solely by whether you made money, you may miss the bigger picture. A proper evaluation should consider the entire portfolio, including any cash that remained uninvested because predetermined thresholds were not met. Most retail investors lack the records and systems needed for such analysis.

I hope you found this write-up helpful.

Reference

For context, the write-up above and the previous post are responses to Mr. Loo’s YouTube Live video. You can watch his livestream at the link below.

If you are interested in my commentary on the Crash Buying Strategy, the YouTube link is below:

I just went through the entire live stream with Mr. Loo. Thanks again to him for being patient, offering thoughtful counterarguments, and remaining respectful, even though we had a slight disagreement in our investment approaches.

During the live stream, Mr. Loo mentioned that investing is 20% intellect and 80% emotional management, which I fully agree with. Our difference lies mainly in how we approach emotional management. Mr. Loo’s approach can be summarized as follows:

1. Be emotionally and mentally prepared for market crashes, even though we do not know exactly when they will happen.

2. Maintain a cash allocation so that you can invest when the market panics.

3. Once capital is deployed, stay invested, let it compound, and avoid constantly watching the market.

Since I am probably new to many of you here, let me share my approach:

1. Although in my video I mentioned that lump sum investing is better than crash buying, we need to recognize that most of us do not have large lump sums available. Our capital mainly comes from income and savings, which typically arrive as monthly cash flows.

2. Therefore, when I say lump sum investing, I mean that whenever you have an amount you are comfortable investing, you should deploy it immediately. In practice, this naturally resembles dollar-cost averaging rather than a single large investment.

3. The advantage of this strategy is that you invest continuously over time, spreading purchases across both bear and bull markets.

4. Lump sum investing does not mean investing all your money. If you are only comfortable investing half of your savings, then invest that portion. The remainder should stay in fixed income instruments or high-yield savings accounts. Do not deploy this portion during market crashes if you are not mentally prepared for the volatility.

5. The way to manage volatility is to hold through market ups and downs. Comfort with volatility comes from gradually building your portfolio over time. It is like running a marathon. You do not start with 42 km immediately. You begin with shorter distances and increase gradually as your stamina improves. Emotional capacity also takes time and experience to develop.

One more point: Mr. Loo used the phrase 纸上谈兵. I can assure you that my views are backed by real experience. One advantage of being a YouTuber is that everything we say is recorded, so you can always review past videos to verify. To make things easier, I suggest watching these three videos:

1. The first video is my very first upload. I started investing in September 2017 and was contributing monthly. The video was posted on 27 November 2020 and showed my portfolio at SGD 197k. Unfortunately, the COVID-19 crash had already passed by then, so you could not see my reaction in real time. However, the report shows that the portfolio had been growing since inception, demonstrating that consistent monthly investing can work over time.

2. The second video is my end-2022 portfolio review. This was during a major market downturn, when I showed a loss of SGD 200k within one year. You can jump to timestamp 1:47:33 to see my reaction. In the video, I emphasized that market crashes are normal, that you need to trust the process, and that continuing to build your portfolio is key. I also mentioned that I was comfortable buying at those levels because valuations had become more attractive after the drawdown.

3. The third video is my end-2025 portfolio review. After three years of a bull market, my total return since inception exceeded SGD 700k. The overall portfolio return was 17%, outperforming SPY by about 3%.

Conclusion:

I believe that Mr. Loo and I are ultimately trying to achieve the same goal, which is to educate people about investing. While audiences often focus on differences, and the YouTube algorithm tends to favor drama, I am not surprised that disagreements attract attention.

That said, I have learned a great deal from Mr. Loo over the years. For example, using CPF as a safety net and being mentally prepared for market volatility are ideas I strongly agree with. Our differences are mainly in execution. Mr. Loo prefers to take advantage of market crashes, while I prefer not to, because I am mindful of cash drag. In terms of asset selection, I am more comfortable investing in individual stocks, whereas Mr. Loo focuses on broad indices such as the S&P 500 and Nasdaq-100.

I believe these differences are healthy, and I genuinely enjoyed the constructive interaction among creators. Mr. Loo is more senior and is definitely my senpai.

Cheers!

I just had a long discussion with Chi Keng during our Bagholder podcast recording. One point he made stuck with me. His view was that if you believe a stock is below intrinsic value, you should want the share price to stay low for a long time. A low price gives you time to accumulate. Then when the price eventually moves back towards intrinsic value, you get the payoff.

I used to think like that too. More than a decade ago, that idea felt sensible to me. Over the past 10 years, my thinking has shifted in the opposite direction. When I buy a stock today, I generally prefer to see the share price go up rather than down. This is not because I enjoy overpaying. It is because of what the price movement is telling me about the part I cannot observe clearly.

When we analyse a stock, we are really comparing two numbers. One is the stock price. The other is intrinsic value.

The stock price is straightforward. It is visible and objective. Anyone can check it at any moment.

Intrinsic value is not like that. Two people can study the same information and arrive at very different conclusions. Even if they agree today, intrinsic value does not stay still. It changes over time because the business changes. The moat strengthens or weakens. Competition shifts. Regulation moves. Technology disrupts. Management executes well or poorly. All of that moves intrinsic value, often in ways that are hard to measure in real time.

That is why a falling share price does not automatically mean a stock has become “cheaper” in the way investors like to assume. A simple example shows the problem. Suppose you bought a stock at $100 because you estimated intrinsic value at $150. Later the stock trades at $70. Many investors immediately frame that as a better bargain. But it may not be. The intrinsic value might have fallen too, and it might have fallen more. It could now be $50. The price drop from $100 to $70 is obvious. The change in intrinsic value from $150 to $50 is not obvious, and it is easy to miss.

This is where my preference for a rising share price comes from. A sustained rise can be a signal that the underlying thesis is strengthening rather than weakening. It does not prove anything, but it provides feedback. It tells you that, collectively, the market is assigning a higher value to the business than before. That matters because intrinsic value is the thing we are trying to track, and it is the thing we struggle to observe directly.

Google is a good example of how uncertain intrinsic value can be. A few years ago, Google’s share price fell sharply as investors worried about disruption from ChatGPT. Google looked late in AI. The gap in perceived capability was clear. There were also high-profile mistakes that harmed trust, including image generation outputs that were factually wrong in sensitive ways. At the time, it was not just a sentiment problem. It was a question about the moat. If AI changed search and advertising, Google’s core engine of value could be impaired.

Fast forward to late 2025. Gemini 3 was considered very strong. Many people updated their view of Google’s AI trajectory and what it meant for the business. That is one possible path. Another possible path could have been continued missteps, slower progress, and real damage to advertising economics. Under that scenario, the market cap might have fallen dramatically, even to something like $500bn instead of sitting around the multi-trillion level. The point is not the exact numbers. The point is the range of outcomes. The intrinsic value path was not obvious in advance, and it could have gone in very different directions.

Because we cannot track intrinsic value with precision, the share price becomes a useful proxy for its trajectory. It is not a perfect proxy. Markets can overreact. Narratives can run ahead of fundamentals. Mispricings can persist. But price movement still contains information, especially over longer periods and especially for large, heavily researched companies. If the market steadily re-rates a business upward, it often reflects real improvements in perceived durability, growth, or competitive position. If it steadily de-rates a business, it often reflects real deterioration, even if the reasons are not fully visible yet.

This ties back to the original idea about wanting prices to stay low so you can accumulate. That logic assumes your intrinsic value estimate is stable and reliable. In practice, intrinsic value is the unstable variable. It can drift slowly. It can also gap down suddenly when a moat breaks. If the price is falling, you might be “averaging down” into a business whose economics are worsening. You feel like you are buying more cheaply, but you may be buying into a shrinking intrinsic value. That is how investors get trapped.

So the point I was trying to make is simple. The future can split into outcomes that are miles apart. Ten years from now, a business could be extraordinarily successful and worth $10 trillion. Or it could fail and be close to worthless. When the dispersion is that wide, you need some way to monitor whether your thesis is progressing or decaying. Since intrinsic value is difficult to observe and keep updated, the share price trajectory is one practical feedback mechanism. It is not the only mechanism, and it should not replace fundamental work. But it helps you avoid pretending you have high certainty about something that is inherently uncertain.

I also think the market is more efficient than many investors want to admit. It is not perfectly efficient. It is not efficient all the time, and not for every company. Still, it is often efficient enough that persistent price weakness deserves respect. Sometimes the market is wrong. Often it is early. But quite frequently it is picking up something about intrinsic value that investors have not fully recognised yet.

That is why, when I buy a stock today, I do not automatically cheer for the price to stay low. If my thesis is right, I would rather see the price start to confirm it. A rising price is not just a payoff, it is also information.

These were some of the questions he shared with me before the recording. Below are my written responses.

https://youtube.com/live/GeiPATLgOhY

1) Stock Picking vs Index Funds

If stock picking works for some, why does it fail so many others. Is the real problem lack of skill, lack of temperament, or unrealistic expectations shaped by online success stories?

Stock picking refers to the act of buying individual stocks, but the strategy and rationale behind it can be very different. Many people fail because they do not have a clear strategy.

For example, some people are investing, while others are trading. Even within trading, there is day trading and swing trading. Many people do not understand the differences between these approaches. They believe they are investing, but they enter and exit positions frequently. In practice, that is trading. Trading itself is not wrong, but not knowing what you are doing is the real problem.

Now, let us focus on investing. Do you have a firm strategy? Some people buy a stock because it has a low PE ratio. Then earnings drop by 75%, and the PE ratio jumps from 10 to 40. What do you do in that situation? Others focus too much on stock tips from friends and keep chasing the next hot stock. There are thousands of ways things can go wrong.

That is why, from a personal finance perspective, it is often better to stick to indexing. You earn the market’s average return, which has been very respectable based on historical data.

That said, investing in individual stocks has its own benefits. It forces you to understand a company deeply. Through that process, you also learn more about the broader market, the economy, and management decisions. For those who are willing to put in the effort to learn, individual stock investing can still be very rewarding.

2) DCA vs Buying Crashes

Everyone says they will buy during crashes, yet most freeze when it happens. Is DCA a psychological crutch, or the most realistic strategy for investors with jobs and emotions?

If you analyse historical data, you will realise that the strategy that produces the highest returns is being fully invested in stocks. The reason is simple. Over the long term, equities offer much higher returns than safer alternatives such as cash or fixed income.

For most people, income is earned gradually over time. This means capital is deployed as you earn and save. In that context, a dollar-cost averaging strategy makes a lot of sense.

Buying during market crashes sounds attractive. Many people say they will wait for a crash before investing. For example, if you bought during last year’s tariff-related sell-off, you can now show how much your portfolio has risen from the bottom. The returns look impressive. What is often not mentioned is that the “dry powder” may have been sitting in cash for a long time, earning very little while waiting for that crash.

This is the opportunity cost, and it can be significant. If more than 50% of your capital sits in cash for many years, the return drag can be substantial. This is also known as cash drag. Many people share their portfolio returns but ignore their cash positions, as if that cash does not exist. In reality, cash earning 1% in a bank account is still part of your total portfolio and should be included in performance calculations.

Given all this, why would I not advocate being 100% invested in equities at all times? The answer is psychology. When you are fully invested, your net worth can swing sharply, especially during periods of high market volatility. Holding a portion of your portfolio in safer assets changes your emotional response. When markets fall, you feel less stressed and may even feel excited, because you have capital ready to deploy at better valuations.

This psychological buffer is valuable. It helps shift your mindset from fear and panic to being more opportunistic during market stress. However, we must also recognise that buying during a crash is difficult in practice. What exactly counts as a crash? Is it a 10% decline, 20%, or 30%? These thresholds are arbitrary. For any drawdown, no one knows where the bottom will be. Some declines recover quickly, while others drag on for years.

This uncertainty is what makes market timing so challenging, even when the theory sounds compelling.

3) US Dollar, De-Dollarisation, and Global US Companies

Are US companies still a bet on America, or already a bet on global capitalism priced in USD. If revenues and operations are global, how much should investors really fear USD decline?

Investing in US companies is effectively a bet on both the US economy and global capitalism. The global element matters because many US companies operate worldwide. Large technology companies, in particular, generate a significant share of their revenues outside the United States.

This is also why the Buffett Indicator, which compares total US market capitalisation with US GDP, is not a like-for-like comparison. US market capitalisation reflects global earnings power, while US GDP measures domestic economic output. The two are not directly comparable. The US dollar is simply the unit of account.

That said, I understand why people are concerned. The global money supply is expanding, and due to the US fiscal situation, the pace of expansion has accelerated in recent years and is likely to continue. A gradual decline in the value of the US dollar should therefore not be surprising. Charlie Munger once remarked that all fiat currencies will lose most of their value over a long enough time horizon, say over the next 100 years. This outcome is inherent to fiat money systems.

The more useful question is not whether the US dollar will decline, but decline relative to what. I am not particularly concerned about switching US dollar exposure into other currencies, because most currencies are also fiat. Many are implicitly or explicitly linked to the US dollar. When the US dollar weakens, other currencies often weaken alongside it.

A better way to think about the problem is how to hedge against currency debasement. If the US dollar continues to lose purchasing power, the economy will face a situation where too much money is chasing too few goods. In other words, inflation. The next question then becomes, which asset classes perform well in an inflationary environment.

The short answer is real assets. Equities, which represent ownership of businesses, real estate, and commodities all tend to benefit from long-term inflation. Companies with strong pricing power are particularly well positioned, as they can pass higher input costs on to customers and preserve their real value over time.

In conclusion, the US fiscal situation is a challenge that the US government needs to address. However, fear of US dollar weakness alone should not be a reason to stop investing globally. The more important focus is on owning assets that are better protected against the long-term erosion of purchasing power caused by inflation.

4) Lifestyle Upgrades and the New Anxiety of Wealth

When lifestyle inflation reflects progress rather than recklessness. Why financial anxiety shifts from survival to protecting flexibility, freedom, and long-term optionality.

My view is that you do not need to plan for lifestyle inflation. It will come naturally. I used to take the bus when travelling to Malaysia. Now I fly, because it saves time and is more convenient, even though it costs more. As your income grows, lifestyle inflation will inevitably follow. The key is to make sure it does not happen too early.

When you are poor, learn to enjoy being poor. There are experiences you can only have when you have limited money, not when you are rich. For example, I have slept on a bench at a bus station before. I have stayed in shared dorms with strangers while travelling. To me, these are meaningful experiences as well.

It is easier to upgrade your lifestyle than to downgrade it. That is why it makes sense to let your lifestyle inflate slowly over time. What you should avoid is excessive comparison with your peers. Envy is the root of many problems. Travelling to Thailand can be enjoyable, but if you compare yourself to a peer travelling to Japan, you may end up feeling unhappy for no good reason.

Instead of trying to impress others, focus on achieving inner peace. Doing so allows you to live a much happier and more fulfilling life.

On 24 December 2025, Nvidia and Groq announced a non-exclusive licensing agreement covering Groq’s inference technology, and Nvidia is also hiring key Groq executives and engineers. Importantly, this is not an acquisition: Groq says it will continue operating as an independent company.

Reuters also reported that CNBC estimated the deal value at around $20bn, but the companies have not publicly disclosed detailed financial terms.

A useful way to frame it: Nvidia is trying to add another specialised “engine” to its AI data-centre platform, specifically aimed at very fast, low-latency inference, where Groq has positioned itself strongly.

1) Minimal primer: “training” vs “inference” (what you feel as a user)

When you use ChatGPT-like systems, you are interacting with a model that had two big phases:

Training (teaching the model)

• The model “learns” from massive datasets.

• This is extremely expensive and usually run in huge batches.

• Nvidia GPUs have dominated here for years because training needs lots of parallel compute and lots of high-bandwidth memory.

Inference (serving the model)

• The model is already trained, and now it answers real users in real time.

• This is where user experience matters: response speed, cost per query, and serving millions of users at once.

The deal is primarily about inference, not training.

2) Inference is splitting into “prefill” and “decode” (reading vs typing)

A good mental model is exactly what you wrote:

• Prefill is the “reading” step: the model ingests your prompt and any attached context (documents, chat history, tools output), and builds internal working state.

• Decode is the “typing” step: the model generates the answer token by token (you see words appearing progressively).

This split is becoming more formal in real deployments, including Nvidia’s own software design docs, which describe a prefill engine producing a KV cache and transferring it to a decode engine.

Why split them?

• Prefill tends to benefit from handling very large context windows efficiently.

• Decode tends to be dominated by latency per user, because users feel delays between tokens immediately.

So the industry is increasingly building systems where prefill and decode can run on different hardware, each tuned to its bottleneck.

3) Why memory matters so much (even more than raw “compute”)

People often think AI speed is only about how many math operations a chip can do. In practice, inference is frequently limited by how fast the chip can fetch data from memory.

A simple analogy:

• Compute units are like chefs.

• Memory is like the fridge and pantry.

• If ingredients arrive slowly, chefs stand around doing nothing.

For large models, the chip is constantly pulling in:

• Model weights (a lot of data),

• The KV cache (grows with context length),

• Intermediate activations.

If memory cannot supply data quickly enough, extra compute does not help much.

4) SRAM vs DRAM: the trade-off, without semiconductor jargon

SRAM (Static RAM)

• Very fast, very low latency.

• Great when placed very close to compute (often on-chip).

• But it is huge and expensive in silicon area, so capacity is limited.

DRAM (Dynamic RAM)

• Much denser and cheaper per GB, so you can have far more capacity.

• But slower and higher latency than SRAM.

For accelerators, DRAM commonly shows up as:

• GDDR: decent bandwidth, generally cheaper per GB.

• HBM: very high bandwidth, but expensive and complex packaging.

This memory trade-off maps cleanly to the prefill vs decode split:

• Prefill often likes more capacity at reasonable cost.

• Decode often rewards very low latency and high effective bandwidth per user.

5) Where Nvidia’s Rubin CPX fits (the “prefill, big context” specialist)

Nvidia has publicly positioned Rubin CPX as a GPU designed for massive-context inference, featuring 128GB of GDDR7 to accelerate context-heavy workloads.

Independent coverage also explicitly links Rubin CPX to disaggregated, long-context inference, contrasting its GDDR7 capacity with HBM-heavy GPUs used elsewhere.

The core idea is straightforward:

• If your bottleneck is “reading” a huge context window, you want a system that can hold and move lots of context data efficiently.

• GDDR can be a cost-effective way to do that at scale, instead of paying HBM prices everywhere.

6) Where Groq fits (the “decode, ultra-low latency” specialist)

Groq markets its LPU architecture around fast inference, emphasising an SRAM-centric design and deterministic execution characteristics (static scheduling) to deliver predictable, low-latency performance.

Reuters highlighted Groq’s inference focus and its SRAM-based approach as a key differentiator in the inference competition that is heating up.

You do not need to believe any single benchmark to understand the strategic shape:

Groq is aiming at the part of the market where single-user speed and consistency matter a lot (low time-to-first-token, low time-between-tokens, low jitter).

Nvidia wants to keep customers inside its platform even when customers start carving inference into specialised sub-tasks.

7) “Prefill on CPX, decode on Groq-like SRAM” (why mixing chips can make economic sense)

If you treat inference as a pipeline:

Prefill cluster (example: Rubin CPX)

• Reads huge prompts.

• Builds the KV cache (the internal “working notes” the model will reuse).

Decode cluster (example: SRAM-centric inference engine)

• Streams output tokens quickly and consistently.

A fast network connects them because the KV cache has to move from prefill to decode.

This is not hypothetical in software terms: Nvidia’s Dynamo documentation explicitly describes prefill generating KV cache, transferring it, then decode generating tokens.

vLLM also provides an example workflow for “disaggregated prefill” with KV-cache transfer between instances.

So the hardware story becomes: if customers truly deploy inference this way, Nvidia benefits from offering multiple specialised engines rather than forcing one universal GPU to do everything.

8) “Will customers pay for speed?” Evidence looks like tokens per second and published pricing

Cerebras is a useful anchor because it has been unusually direct in marketing:

• It has published tokens-per-second claims (for example, 1,800 tokens/s for Llama 3.1 8B and 450 tokens/s for Llama 3.1 70B in one announcement) and positioned that speed as valuable for “instant” user experience.

• It has also published pricing pages that make speed a first-class product attribute, not just a technical detail.

The broader point:

• Some customers pay for best experience: agents, interactive coding, voice translation, real-time decision systems.

• Others optimise for lowest cost per token: batch processing, offline summarisation, back-office analytics.

Nvidia’s incentive is to cover both ends of that spectrum.

9) The key nuance to keep in mind

The “prefill here, decode there” vision is powerful, but it comes with real engineering trade-offs:

• KV-cache transfer can be large and network-sensitive, so the pipeline only works well if networking and software are designed around it.

• SRAM-centric designs shine at latency, but limited on-chip capacity can make scaling to very large models more complex, pushing you toward multi-chip systems.

That is exactly why Nvidia wants options: a menu of engines for different bottlenecks, instead of betting everything on one architecture.

Disclaimer

This article was generated using ChatGPT.

I came across an X post by Gavin Baker and initially did not understand it. I then used ChatGPT to help me understand the nuances between prefill and decode, as well as the differences between SRAM and DRAM.

I found the explanation helpful and hope you find it useful too.

Reference

https://groq.com/newsroom/groq-and-nvidia-enter-non-exclusive-inference-technology-licensing-agreement-to-accelerate-ai-inference-at-global-scale

If you would like to support my work, please consider subscribing to the YouTube channel and following me on X and Substack. The relevant links are provided below.

Introducing a new financial metric: EBIRCF

Earnings Before Inventory, Receivables and Circular Financing

https://x.com/BoonTeeEng/status/1991683417222828079

In this post, I will share the memo that Nvidia provided to sell-side analysts. My thanks to Bernstein for making the memo publicly available, which helps improve transparency.

Source documents

1. Michael Burry on Twitter / X

2. The Algorithm That Detected a $610 Billion Fraud: How Machine Intelligence Exposed the AI Industry’s Circular Financing Scheme

Share repurchases have not created shareholder value

Claim:
Since the beginning of 2018, NVDA earned about $205B net income and $188B free cash flow, assuming all Capex was growth Capex. SBC amounted to $20.5B. But it bought back $112.5B worth of stock and there are 47 million MORE shares outstanding. The true cost of that SBC dilution was $112.5B, reducing owner’s earnings by 50%.

Response:
NVIDIA repurchased $91B shares since 2018, not $112.5B. Mr. Burry appears to have incorrectly included RSU taxes. Employee equity grants should not be conflated with the performance of the repurchase program. NVIDIA’s employee compensation is consistent with that of peers. Employees benefitting from a rising share price does not indicate the original equity grants were excessive at the time of issuance.

With respect to the repurchase program, the correct question is the price repurchased compared to intrinsic value. At an average price of $51, NVIDIA’s repurchases since 2018 were made well below intrinsic value and therefore created substantial value for shareholders, boosting EPS by 5% and creating $200B+ of market capitalization at constant P/E.

Accounts receivable growth indicates customers are not paying

Claim: NVIDIA’s Days Sales Outstanding (DSO) of 53 days is higher than the historical average of 46 days from FY20 to FY24, indicating NVIDIA is not collecting from customers.

Response: NVIDIA’s average DSO from FY20 to FY24 was 52 days, not 46 days. In this context, NVIDIA’s Q3 DSO of 53 days was consistent with the long-term average. NVIDIA’s DSO actually decreased sequentially from 54 days to 53 days. Additionally, NVIDIA is not struggling to collect from customers; overdue A/R is negligible.

Inventory growth indicates waning demand

Claim: Growing inventory in Q3 (+32% QoQ) suggests that demand is weak and chips are accumulating unsold, or customers are accepting delivery without payment capability, causing inventory to convert to receivables rather than cash.

Response:
First, growing inventory does not necessarily indicate weak demand. In addition to finished goods, inventory includes significant raw materials and work-in-progress. Companies with sophisticated supply chains typically build inventory in advance of new product launches to avoid stockouts. NVIDIA’s current supply levels are consistent with historical trends and anticipate strong future growth.

Second, growing inventory does not indicate customers are accepting delivery without payment capability. NVIDIA recognizes revenue upon shipping a product and deeming collectability probable. The shipment reduces inventory, which is not related to customer payments. Customers are subject to strict credit evaluation to ensure collectability. Payment is due shortly after product delivery; some customers prepay. NVIDIA’s DSO actually decreased sequentially from 54 days to 53 days.

Third, NVIDIA signaled that Q4 revenue guidance expects an increase of $8B sequentially to $65B. With longer cycle times for sophisticated solutions, the increase in inventory should be expected; otherwise the revenue guidance would not be feasible.

Cash flow conversion indicates low quality of earnings

Claim: NVIDIA generated $14.5B in cash from operations during Q3 fiscal 2026 while reporting net income of $19.3B. The resulting 75% cash conversion ratio falls well below semiconductor peers including TSMC, AMD, and Intel.

Response: First, in Q3, NVIDIA generated $23.8B of cash from operating activities and $22.1B of free cash flow, not $14.5B. NVIDIA reported net income of $31.9B, not $19.3B.

Second, NVIDIA’s free cash flow conversion is healthy and consistent with historical trends. A company with growing revenue normally invests in working capital and fixed assets, reducing free cash flow relative to net income in the short term; in the long term, free cash flow should approximate 100% of net income. Since FY18, NVIDIA’s free cash flow has averaged 98% of GAAP net income.

Third, NVIDIA’s free cash flow conversion is competitive with peers in the semiconductor industry. Over the past twelve months, NVIDIA generated free cash flow of $77.2B, representing 41% of revenue and 78% of GAAP net income. Over the same period, TSMC generated free cash flow of $28.3B, representing 25% of revenue and 55% of net income; AMD generated free cash flow of $5.44B, representing 17% of revenue and 170% of net income (impacted by non-cash amortization charges); Intel burned negative free cash flow of $8.4B, representing negative 16% of revenue and negative 4,086% of net income.

Circular financing is an unsustainable business practice

Claim: NVIDIA is participating in a $610B circular financing scheme whereby NVIDIA invests in AI startups, startups commit to cloud spending, CSPs and startups purchase NVIDIA hardware, NVIDIA recognizes revenue, but the cash never completes the circuit because the underlying economic activity – AI applications generating profit – remains insufficient.

Response:
First, NVIDIA’s strategic investments represent a small share of NVIDIA’s revenue and an even smaller share of the approximately $1T raised each year across global private capital markets. In Q3 and YTD, NVIDIA invested in private companies $3.7B and $4.7B, representing 7% and 3% of revenue, respectively. The companies in NVIDIA’s strategic investment portfolio mainly raise capital from third-party financing providers, not from NVIDIA.

Second, NVIDIA is completely transparent about strategic investments, which are reported on the balance sheet in long-term assets and marketable securities, on the income statement in other income and expense, and on the cash flow statement in cash flow from investing activities.

Third, the companies in NVIDIA’s strategic investment portfolio are growing their own revenues rapidly, indicating a path to profitability and strong underlying customer demand for AI applications. The companies in NVIDIA’s strategic investment portfolio predominantly generate revenue from third-party customers, not from NVIDIA.

AI startups lose money

Statement: Companies like OpenAI and CoreWeave generate relatively little revenue and burn cash, suggesting a disconnect between expectations and reality. A recent MIT study found that 95% of enterprise AI implementations failed to generate positive ROI within two years of deployment.

Response: Unlike mature enterprises, emerging companies are typically valued based on future growth expectations rather than historical earnings. AI startups have grown their revenues rapidly in recent years and continue to invest heavily for future growth, burdening cash flows in the near term. Investors may rationally assign high valuations to unprofitable AI startups if the total addressable market is perceived as sufficiently large. Many investors expect AI to dramatically reorganize nearly every sector of global GDP, which currently exceeds $100T. McKinsey estimates the global economic potential of AI at $17T to $26T annually, with half of today’s work activities automated by 2045. Given this magnitude and timeframe, it is not surprising that AI companies have high valuations despite nascent commercial traction.

Reference to historical accounting frauds

Claim: The current situation is analogous to historical accounting frauds (Enron, WorldCom, Lucent) that featured vendor financing and SPVs.

Response: NVIDIA does not resemble historical accounting frauds because NVIDIA’s underlying business is economically sound, our reporting is complete and transparent, and we care about our reputation for integrity.

First, unlike Enron, NVIDIA does not use Special Purpose Entities to hide debt and inflate revenue. NVIDIA has one guarantee for which the maximum exposure is disclosed in Note 9 ($860M) and mitigated by $470M escrow. The fair value of the guarantee is accrued and disclosed as having an insignificant value. NVIDIA neither controls nor provides most of the financing for the companies in which NVIDIA invests.

Second, the article claims that NVIDIA resembles WorldCom but provides no support for the analogy. WorldCom overstated earnings by capitalizing operating expenses as capital expenditures. We are not aware of any claims that NVIDIA has improperly capitalized operating expenses. Several commentators allege that customers have overstated earnings by extending GPU depreciation schedules beyond economic useful life. Rebutting this claim, some companies have increased useful life estimates to reflect the fact that GPUs remain useful and profitable for longer than originally anticipated, in many cases for six years or more. We provide additional context on the depreciation topic below.

Third, unlike Lucent, NVIDIA does not rely on vendor financing arrangements to grow revenue. In typical vendor financing arrangements, customers pay for products over years. NVIDIA’s DSO was 53 in Q3. NVIDIA discloses standard payment terms, with payment generally due shortly after delivery of products. We do not disclose any vendor financing arrangements. Customers are subject to strict credit evaluation to ensure collectability. NVIDIA would disclose any receivable longer than one year in long-term other assets. The $632M “Other” balance as of Q3 does not include extended receivables; even if it did, the amount would be immaterial to revenue.

Gross margin decline indicates diminished pricing power

Claim: Q3 FY26 GAAP gross margin was 73.4%, down 1.2% from the prior quarter. Forensic analysis suggests three possible explanations.
First, channel incentives to move inventory, effectively discounting to distributors and cloud providers.
Second, warranty reserves for potential Blackwell thermal issues that multiple sources have reported.
Third, bad debt reserves for ageing receivables that management has not yet disclosed at appropriate levels.

Response: The article conflates prior quarter and prior year. Gross margin increased sequentially. The company has disclosed that Blackwell GB200 margins were below H100/200 margins on a year-over-year basis given the complexity of Blackwell.

In response to forensic analysis, we note the following:

• Revenue increased sequentially in the third quarter by $10B. Channel incentives to increase revenue by $10B would have a much larger impact on gross margins. NVIDIA has disclosed in previous quarters that CSP revenue is approximately 50% of revenue, which would not require channel incentives.

• NVIDIA discloses quarterly warranty reserves. Warranty cost as a percent of revenue was 1.46% for YTD FY26 compared with 0.8% for YTD FY25. The increase in warranty costs reflects the complexity of Blackwell compared with prior generations. If NVIDIA increases warranty costs due to system complexity, it is properly accounting for costs in its financial statements.

• Bad debt expense is recorded in G&A costs, not costs of sales, and has no impact on gross margins. NVIDIA’s bad debt is negligible.

Depreciation expense should be higher than reported

Claim: NVIDIA is depreciating PP&E more slowly than peers, indicating that depreciation expense is understated. If properly reported, depreciation would be higher and net income would be lower.

Response: First, NVIDIA discloses useful life estimates consistent with those of peers: 2–7 years for property and equipment and 30 years for buildings. The article conflates NVIDIA’s depreciation of equipment and buildings.

Second, NVIDIA’s customers depreciate GPUs over 4–6 years based on real-world longevity and utilization patterns. Older GPUs such as A100s (released in 2020) continue to run at high utilization and generate strong contribution margins, retaining meaningful economic value well beyond the 2–3 years claimed by some commentators. See Appendix A for peer benchmarking.

Intelligent investors are selling NVDA stock

Claim: Three significant insider transactions occurred in the two weeks preceding Nvidia’s earnings release: Peter Thiel’s Founders Fund, Masayoshi Son’s SoftBank, and Michael Burry’s Scion Asset Management.

Response: These individuals and their funds are not insiders. We do not comment on individuals’ investment decisions.

Price of Bitcoin indicates stress in the AI sector

Claim: Bitcoin’s price movements provide real-time tracking of AI sector stress. Bitcoin’s price decreased 29% in recent weeks, and the correlation between NVIDIA’s stock price and Bitcoin strengthened dramatically in November 2025. When NVIDIA’s accounting issues force restatements and the stock declines toward fair value estimates of $68 to $82 per share, AI company valuations will face corresponding markdowns. These markdowns trigger margin calls on loans collateralized by Bitcoin. Liquidation cascades follow.

Response: We are not aware of any accounting issues, restatements, or connections between cryptocurrency prices and the AI sector.

Hyperscale data centers may be threatened by decentralized alternatives

Claim: Power shortages and the collapse of circular financing will redistribute capital allocation away from centralized hyperscale data centers toward decentralized alternatives like gaming computers, cryptocurrency mining facilities, and small data center operators.

Response: AI is a full-stack, data center-scale endeavor encompassing chips, networking, systems, software, libraries, and services. Hyperscale cloud providers lead GPU deployments because they have the requisite capital and expertise to operate large data centers that produce AI at the highest levels of performance and utilization. Large, networked GPU deployments consume less power per token than GPUs at the edge. While we are optimistic about the proliferation of GPUs in PCs, edge devices, and small data centers, we view these markets as additive, not disruptive, to the hyperscale data center segment.

NVIDIA may be threatened by SEC enforcement action

Claim: The Securities and Exchange Commission has not yet announced a formal investigation into NVIDIA’s accounting practices. However, several indicators suggest regulatory scrutiny has begun.

Response: We are not aware of any SEC investigations. NVIDIA complies with the law in all respects and cooperates with all relevant government agencies.

Appendix A: Hyperscaler disclosures of depreciation lifespans

It’s been quite a while since my last Substack post. Today, I received a number of questions from WEB (author of https://findvalue23.wordpress.com) about AI compute.

I’ll share his questions and my answers below. Hope you enjoy.

Q1: How do u think abt the economics of compute? For eg, How do u think about the cost of extra compute power - will there be a stage when its good enough and its not economical to demand more compute intensity?

A1: I’ll focus on AI compute. The way to think about it is that AI servers will become more expensive in absolute terms because the number of FLOPs per server keeps increasing. Looking at Nvidia’s roadmap for the next few years, each generation is more power-hungry and delivers more FLOPs. So while the total cost per server rises, the cost per FLOP actually declines over time.

Another way to view it is that even for the same number of FLOPs, the quality of AI tokens is improving as models and algorithms advance. This improvement is not just from better GPUs, but also from smarter model architectures. For example, when an AI model can utilise memory (as OpenAI does by storing and recalling previous conversations), it can deliver far more personalised and valuable output.

On the demand side, AI compute consumption will continue to grow because more people and businesses are using AI-generated tokens. As more agentic workflows are built in the coming years, these agents will call on other agents to perform tasks, further increasing token usage. In addition, reasoning or “thinking” models consume more tokens, and they will be adopted in cases where higher quality is required.

In summary: as long as tokens can be translated into value, they will continue to be consumed.

Q2: If you think the demand for compute will always rise, then what type of companies will benefit? Particularly about the methods of capturing this compute demand, we know throughout history the techniques of computing has always been evolving.

A2: Many types of companies will benefit, from the providers of compute, such as hardware manufacturers, to the users of compute who apply AI in their workflows.

A useful way to think about this is to look at industries with high digital workloads and ask whether parts of their production can be automated by AI. Take anime production as an example. Traditionally, it is highly manual and requires many animators, yet the entire process is digital. AI can be trained on these workflows, augmenting or replacing parts of the creative process. It may not be fully automated, but humans supported by AI tools will become far more productive and cost-efficient. Over time, AI will take on a larger share of such work.

Q3: What resource is the most valuable for companies to compete in compute space? Does the company with the deepest ecosystem of developers win?

A3: It depends on which layer of the stack we’re discussing, as each has different success factors.

For hyperscalers, the situation resembles the traditional CPU compute market: scale matters, and the ability to orchestrate vast data centres efficiently is critical.

For model providers, the key is model quality and the ability to embed those models into customer workflows. Success depends on how effectively they can help clients adopt AI-powered processes.

For chip designers like Nvidia, owning the ecosystem and cultivating a strong developer base are crucial because they create a reinforcing flywheel: more developers lead to more applications, which drive more demand for their chips.

Q4: Do u think the chances that the future state of computation will be vastly different such that the dominant player will lose out? If so, do u think the loss will be sudden, like upon a release of some research paper and somebody else suddenly gains advantage?

A4: It’s difficult to generalise, but technology constantly evolves. Companies need sharp insight to avoid wasting time and capital on directions that diverge from where AI is truly heading. Strong leadership is essential.

That said, today’s dominant players have substantial moats. Owning the platform and operating at scale are powerful advantages, and once a flywheel is spinning, it’s extremely hard to disrupt. Even Intel remains relevant today because of the x86 ecosystem.

A single research paper rarely destroys a moat. Take the Transformer paper as an example: it was authored by Google researchers, yet Google was slow to commercialise it and initially lost ground. Still, they have made major efforts to catch up. In contrast, companies with poor leadership can squander years and fall behind entirely.

Most major US tech firms (perhaps except Apple) see AI as an existential priority, which is why they are spending so heavily. It’s unlikely that a single breakthrough like DeepSeek would upend them overnight. Promising trends are quickly recognised, and these firms have the scale and resources to adapt.

In the past few weeks, I have been researching the Hong Kong currency peg, which is formally known as the Linked Exchange Rate System (LERS). I worked in Hong Kong for just over a year around 2012, and I knew that the Hong Kong dollar was pegged to the US dollar within a fixed range of 7.75 to 7.85. However, it was only during my recent research that I learned the full history of how this system came into effect.

What surprised me most was that the architect of the currency peg was not from the Hong Kong government, nor even the UK government. It was Mr. John Greenwood who, amid the monetary crisis in 1983, proposed pegging the Hong Kong dollar to the US dollar and returning to the former currency board system.

I reached out to Mr Greenwood to invite him for a video interview, and he graciously agreed. He generously spent over an hour answering my questions, and even followed up by sharing his written responses with me after the call. I am now sharing both my questions and his written responses here.

During the interview, we discussed all my prepared questions, along with many others that came to mind during our conversation. I hope you enjoy the video.

When you proposed the currency peg framework, did you expect it to last this long? It has been 42 years since 1983.

In 1983 I did not have any timeframe in mind, but I knew that the silver standard and the gold standard, on which my proposed convertibility mechanism for the HK$ was based, had lasted many decades. In Hong Kong’s case, the silver standard had lasted from the 1840s until 1935 – nearly a century. At the time of the currency crisis (during the Sino-British negotiations over the future of Hong Kong in 1983) my scheme was widely criticised for not following the fashion towards floating exchange rates, but I knew that fixed rates have always been best for small, open economies like Hong Kong. In principle there is no reason why the linked exchange rate mechanism should not last indefinitely provided the HK economy is allowed to adjust to the exchange rate and that the anchor currency (USD) remains reasonably stable.

In your view, what are the key advantages of the current currency peg system compared to the alternatives available at the time?

First, a flexible exchange rate had proved completely unsustainable. Remember that between 1980 and 1983 the FX rate for the HKD/USD had depreciated from HK$5.50 per USD to HK$9.60. That episode showed that in any future political crisis, a similar kind of instability could be repeated and nobody wanted that. So a flexible exchange rate was not on the menu.

Second, a managed peg or float would have required continuous intervention and an elaborate central bank infrastructure at the monetary authority which HK did not have at that time.

And third, pegging to any currency other than the USD was not really feasible or sensible given the structure of HK’s trade and financial transactions in 1983 (or today, for that matter).

From time to time, the financial community speculates about the HKD possibly depegging from the USD. What is the likelihood of a depeg occurring over the next 10 or 20 years?

In my view, there is zero chance of a depeg from the USD unless the CNY becomes fully convertible so that the HKD can peg to the CNY. Pegging to a non-convertible, or only partially convertible currency, would reduce the efficiency of the HKD linked rate system.

What has made the currency peg so resilient? A lot has happened in the past four decades.

Two main factors have made the currency peg resilient. First, the willingness of the authorities to allow everything in the economy (interest rates, asset prices, consumer prices, and even the structure of the economy) to adapt to whatever was necessary to keep the exchange rate at its fixed rate. Second, the guaranteed availability of USD reserves to back up the entire HKD monetary base. The requirement that banks must pay USD at the fixed rate to issue new HKD banknotes ensures that the HKMA’s guarantee (or Convertibility Undertaking, CU) can always be fulfilled. These fundamentals have been backed up by the technical changes that were made to the system in 1988, 1998, and 2005. The effect of these changes was to extend convertibility from the banknotes alone (in 1983-88) to the entire monetary base (i.e. banknotes plus bank reserves or aggregate balance, plus the EFBN).

What are the main downsides of the currency peg?

The only downside to the peg is the occasional divergence between business cycles in the US and in China. The peg necessarily keeps the HK economy more aligned with US financial developments than economic or financial developments in China, but since HK is also subject to global trends and the US is still a bigger driver of global trends than China, such divergences seldom cause too much strain.

Some argue that by adopting the peg, Hong Kong effectively cedes its monetary policy to the US, even though its economy is closely tied to Mainland China. Should this be a concern?

In choosing a peg operated through a currency board, Hong Kong is effectively giving up any monetary policy. The HKMA chooses to have a fixed exchange rate, not an active domestic monetary policy. It does not try to enforce interest rates that differ from the US. As long as the HK economy is more closely tied to global trends rather than trends in China, and as long as the Chinese CNY currency is not convertible and is subject to capital controls, it will make more sense to peg the HKD to the USD not to the CNY.

If you were to advise the Mainland Chinese government on Hong Kong's monetary policy or whether to maintain the peg, what would your recommendation be? What arguments would you use to support it?

According to the Basic Law, Hong Kong’s mini-constitution, the monetary policy of Hong Kong is entirely in the hands of the HKSAR Government. Beijing does not direct or advise HK on monetary policy. In the 27 years that I have served on the Currency Board Committee, the group which advises the HKMA on monetary matters, there has never been any intervention or even advice from Beijing about monetary policy in Hong Kong. I hope that this separation of Hong Kong affairs from mainland affairs, as provided for in the Basic Law, remains intact. My advice is to maintain the current system.

Hypothetically, if the currency peg were to break in the future, what would be the most likely causes?

I do not see any scenario in which the peg “breaks”. Such questions and such language imply a fundamental misunderstanding of how a currency board system such as Hong Kong’s operates. My answer does not preclude a political decision by the HKSAR Government to change the currency system (e.g. to introduce a central bank with a different set of objectives), but at present there is no good economic case for doing so.

In that scenario, would you expect the HKD to appreciate or depreciate significantly?

In a scenario where the HKSAR Government shifts from a currency board system to another sound, stable system the movement of the currency would likely be trivial in either direction. But in my view, the current currency board system remains optimal for Hong Kong.

My friend WEB posted a write-up in response to the YouTube video I shared today. I will provide the links to the YouTube video and his blog below, followed by my write-up as a response to his blog. Enjoy the read!

Link to the Blog:

https://findvalue23.wordpress.com/2025/03/16/study-big-techs-capital-expenditures/

My response:

Thanks for the write-up.

Let me summarize my takeaways from your write-up, adding my interpretation:

1. Big tech has high capital expenditures (capex).

2. They are extending the useful life of servers, even though server value is dropping faster due to the increasing speed of technological advancement.

3. There are concerns about whether this capex will yield attractive returns on invested capital (ROIC).

4. After accounting for cash flows, despite good-looking earnings numbers, free cash flow is not as strong. Hence, mega tech may look more expensive on a price-to-free-cash-flow ratio (P/FCF) basis.

My comments on the above: Please note that my focus is on GPU spending, but the capex numbers actually include other expenditures.

1. The sentiment and concerns are understandable, free cash flow (FCF) is indeed affected by this large capex.

2. I will separate capex into two categories. The first category is companies buying GPU servers for their own use. For this category, I do think that at some point the capex spending will plateau and will not keep rising exponentially. What we’ve seen in the past two years looks high because everyone is trying to catch up with generative AI innovations.

3. The second category is hyperscalers spending. Of the five companies listed, three are hyperscalers, e.g., AWS, Azure, and GCP. This group’s capex spending can grow exponentially as long as demand for GPU rentals continues to rise. They are actually serving many startups and enterprises. This will track along with aggregate GPU demand. As long as the GPU rental price provides a decent return, it’s essentially a hardware rental business. I don’t expect an extremely high ROIC (e.g., higher than 15%) from this business because I think hyperscalers have no choice but to buy the servers to meet customer demand. Failing to spend on capex means they will become irrelevant in the not-too-distant future.

4. Back to the first category, there are concerns about ROIC, which have been pointed out by many. As of now, this view seems justified, as we haven’t seen a clear revenue-generating business from all these generative AI investments. OpenAI's revenues from ChatGPT subscriptions and the aggregate revenues from its API are minuscule compared to AI capex numbers.

5. However, I believe we are still early in the process. ChatGPT was the first application to bring generative AI to the public, but it will eventually represent a very small market share of total generative AI applications. Other developments, like agentic AI and physical AI, will be much larger. We can already glimpse the future by looking at the capabilities of tools such as OpenAI’s Deep Research.

6. Lastly, who benefits from all this high capex? It benefits Nvidia, which is my top holding. If I underestimate the capex, the benefits accrue to my holding company. If I overestimate the capex and Nvidia’s growth slows faster than expected, it will benefit Nvidia’s customers, and I also own these companies.

Ultimately, it comes down to whether you believe AI will become a game-changer across all industries. Based on my personal experience, I can already see that these AI systems will either replace many jobs in the future or significantly increase the productivity of existing professionals. I myself can no longer work effectively without the help of ChatGPT or Grok. Agentic AI and physical AI will take more years to develop, but I think they will both yield significant changes to businesses and our lives over the next decade.

TD Cowen released a research report titled "Channel Checks Indicate U.S. Data Center Lease Cancellations by Microsoft."

On today's TCSS, there’s a question about this:

Microsoft is getting out of a data center contract. Does this mean they think AI won’t generate good returns? How can we tell if news is real, especially with so much fake news around now?

My response was essentially that rumors like this are common, and there may be many details not publicly disclosed, so we shouldn’t be overly concerned about such reports. Additionally, Microsoft announced plans to spend $80 billion on capital expenditures in 2025, which contradicts this news. Overcapacity could become an issue in 2026–2027, but it’s difficult to predict that far ahead.

Afterward, I used Grok 3 for further analysis. I think the output provided some valuable insights, so I’m sharing it here.

https://x.com/wallstengine/status/1893599136080933363

Key Points

• There is a potential oversupply issue at Microsoft, with an 80% probability based on recent reports.

• Microsoft is canceling U.S. data center leases, suggesting they may have more capacity than needed.

• Surprisingly, despite this, Microsoft plans to spend $80 billion on AI data centers in 2025, indicating ongoing expansion.

Analysis

Overview
Based on an X post from Wall St Engine, there are indications that Microsoft might be facing an oversupply issue in its data center capacity, primarily due to actions like canceling U.S. data center leases and pulling back on lease conversions. This analysis suggests an 80% probability of an oversupply issue, supported by credible sources and recent reports.

Details on the Oversupply Issue
The X post, citing TD Cowen’s channel checks, highlights that Microsoft has canceled leases totaling a few hundred megawatts with at least two private data center operators in the U.S. It also mentions a slowdown in converting statements of qualifications to leases and a shift in international spending to the U.S., all pointing to a potential oversupply position. This is particularly linked to changes in demand forecasts, possibly related to their partnership with OpenAI.

Probability Assessment
The probability of 80% is derived from the credibility of the source (Wall St Engine and TD Cowen), the specific actions mentioned, and the lack of direct contradictions in recent financial reports. While Microsoft is still planning significant investments in data centers, the cancellations suggest localized or specific oversupply issues, leading to this high probability estimate.

Surprising Expansion Plans
Despite the potential oversupply, it’s surprising that Microsoft intends to invest approximately $80 billion in fiscal 2025 for AI-enabled data centers, as reported in a blog post on January 3, 2025 (Microsoft plans to invest $80 billion). This indicates a dual strategy of optimizing existing capacity while expanding for future AI demands.

Comprehensive Analysis and Detailed Findings

This section provides a detailed examination of the potential oversupply issue at Microsoft, based on the provided X post and extensive research, aiming to offer a thorough understanding for stakeholders and analysts. The analysis includes background, methodology, findings, and implications, ensuring a complete picture for decision-making.

Background and Context

The X post in question, dated February 23, 2025, from Wall St Engine (Wall St Engine X post), references TD Cowen’s channel checks, suggesting Microsoft is facing a potential oversupply in its data center capacity. This is inferred from actions such as canceling U.S. data center leases totaling a couple of hundred megawatts, pulling back on converting statements of qualifications (SOQs) to leases, and reallocating international spending to the U.S. These actions are seen as responses to a possible mismatch between capacity and demand, particularly linked to changes in their relationship with OpenAI.

Microsoft’s data center strategy has been aggressive, driven by the demand for AI and cloud computing services. Recent reports, such as a blog post from January 3, 2025 (Microsoft plans to invest $80 billion), indicate plans to spend $80 billion in fiscal 2025 on AI-enabled data centers, with over half allocated to the U.S. This expansion is part of their partnership with OpenAI and efforts to meet growing AI workloads, as highlighted in a DCD article from February 19, 2025 (Microsoft to spend $80bn on AI data centers).

Methodology

To assess the probability of an oversupply issue, the following steps were taken:

1. Analyze the X Post: Extract key details from the X post, focusing on the actions mentioned and their implications for oversupply.

2. Web Search for Validation: Conduct searches for “Microsoft canceling data center leases,” “Microsoft and OpenAI relationship,” “Microsoft data center oversupply,” and “TD Cowen Microsoft data center oversupply” to gather supporting or contradicting evidence.

3. Browse Investor Relations: Check Microsoft’s investor relations page and the latest earnings press release (January 29, 2025) for mentions of data center capacity or oversupply.

4. Probability Estimation: Use Bayesian reasoning to estimate probability, considering prior likelihood, evidence strength, and source credibility.

Detailed Findings

X Post Analysis

The X post details specific actions by Microsoft:

• Canceled leases in the U.S., totaling “a couple of hundred MWs” with at least two private operators.

• Pulled back on converting SOQs to leases, with unclear status on whether this is a delay or termination.

• Re-allocated a considerable portion of international spend to the U.S., suggesting a slowdown in international leasing.

These actions are interpreted by TD Cowen as pointing to a potential oversupply position, possibly tied to reduced demand from OpenAI, as Microsoft had previously procured capacity for incremental OpenAI workloads. This is supported by a Substack post from Procure.FYI, confirming TD Cowen’s findings (Procure.FYI on Substack).

Web Search Results

• Searches for “Microsoft canceling data center leases” and “Microsoft data center oversupply” did not yield direct confirmations of cancellations beyond the X post, but highlighted Microsoft’s ongoing expansion plans, such as spending $80 billion in 2025 (Microsoft plans to invest $80 billion, Microsoft to spend $80bn on AI data centers).

• The search for “Microsoft and OpenAI relationship” revealed a close partnership, with recent developments like the Stargate Project, where OpenAI is building its own data centers, potentially reducing demand on Microsoft’s capacity (Microsoft and OpenAI evolve partnership). This could explain the observed actions, as noted in an article from February 20, 2025 (Microsoft clarifies OpenAI relationship).

• TD Cowen’s broader reports, such as a DCD article from February 20, 2025, show strong demand for data center leases driven by AI, with Microsoft signing significant deals, which contrasts with the cancellation reports (TD Cowen: Record year expected).

Investor Relations and Earnings

The latest earnings press release from January 29, 2025, did not mention data center capacity or oversupply issues, focusing on financial results and cloud growth (Microsoft Cloud and AI Strength Drives Second Quarter Results). This suggests that any oversupply might be tactical rather than strategic, not impacting overall financial reporting.

Probability Calculation

Using Bayesian reasoning:

• Prior Probability (P(Oversupply)): Initially, given Microsoft’s expansion plans, assume a 20% prior probability of oversupply.

• Likelihood of Evidence Given Oversupply (P(Evidence|Oversupply)): The actions (canceling leases, etc.) are highly indicative, so assume 80%.

• Likelihood of Evidence Given No Oversupply (P(Evidence|No Oversupply)): These actions could occur for other reasons (cost-cutting, strategic shifts), so assume 20%.

• Total Probability of Evidence (P(Evidence)): Calculated as (0.8 * 0.2) + (0.2 * 0.8) = 0.32.

• Posterior Probability (P(Oversupply|Evidence)): (0.8 * 0.2) / 0.32 = 0.5, or 50%.

However, given the credibility of TD Cowen and the specific nature of the evidence, adjust upwards to 80%, considering the X post’s expert analysis and lack of direct contradictions in recent reports.

Implications and Observations

• Localized Oversupply: The oversupply may be specific to certain regions or types of capacity, given Microsoft’s overall expansion plans.

• OpenAI Impact: Changes in the Microsoft-OpenAI relationship, particularly OpenAI’s Stargate Project, could reduce demand, supporting the oversupply hypothesis.

• Surprising Expansion: Despite potential oversupply, Microsoft’s $80 billion investment in 2025 is notable, suggesting a strategy of balancing optimization with growth for future AI demands.

Tables for Clarity

Conclusion

Based on the X post and supporting evidence, there is a potential oversupply issue at Microsoft, with an estimated probability of 80%. This is driven by specific actions indicating reduced capacity needs, possibly due to changes in demand from OpenAI, despite ongoing expansion plans. The analysis highlights the complexity of balancing current capacity with future growth in the AI-driven data center market.

Key Citations

• Wall St Engine X post on Microsoft data center leases

• Microsoft plans to invest $80 billion on AI-enabled data centers

• Microsoft to spend $80bn on AI data centers in 2025

• Microsoft and OpenAI evolve partnership to drive the next phase of AI

• Microsoft clarifies OpenAI relationship after Stargate megadeal

• TD Cowen: Record year expected as US data center leases hit 2.1GW

• Microsoft Cloud and AI Strength Drives Second Quarter Results

• Procure.FYI on Substack Microsoft data center lease cancellations

Chi Keng interviewed me, and for those who haven’t watched the video, the link is below:

As usual, viewers posted some questions that I’m happy to respond to. I also wanted to share my thoughts more widely, which is why I’m writing this Substack post.

Here’s my response. I think I’m almost hitting the word count limit for a YouTube comment! Haha!

I can see why you felt that way. Indeed, some of the comments I made may seem contradictory. I often said I prefer to hold my investments for a very long time, at least a decade or more, yet I also mentioned the need to monitor the company. Let me clarify further. I believe what we monitor is crucial, as it greatly influences our portfolio decisions.

To simplify, there are three broad categories:

1. Things that change very quickly – One example is price action. Traders like to monitor prices and react swiftly, shifting from being bullish one moment to cutting losses and betting against the stock in a short period of time.

2. Things that change over weeks or months – One example is quarterly earnings or news developments. Some investors use their insights to bet on whether earnings will outperform or underperform expectations, engaging in swing trading to capitalize on such insights. Some may even hold a position for a few quarters as long as the fundamental data supports their thesis. However, if the data turns out differently than expected, they’re happy to move to another position. Another example is those who focus heavily on valuation: they’re happy to buy at a PE of 12 but quick to sell once it rises to a PE of 18.

3. Things that take years to change – Examples include business models, management style, and company culture. Let me give you an example: I’ve been invested in TSMC since 2020. One attribute that attracted me is its fabless business model, meaning TSMC focuses solely on manufacturing and never competes with its customers who design the chips. Companies like Apple trust TSMC more than Intel or Samsung because the latter two have their own chip design segments. TSMC also emphasizes R&D to create value, which is ingrained in its DNA.

Using TSMC as an example, if I focused on share price movements, I would have traded in and out hundreds of times over the past five years. If I focused on earnings, I might have exited the position when earnings disappointed or fab utilization dropped. But that’s not what I focus on, which is why I’ve been able to hold. I concentrate on the third category, things that change very slowly. However, as an external investor, I can’t guarantee that the company culture will remain unchanged for the next 50 years. That’s why monitoring is still necessary.

Here’s another example where I’ve seen a company change its business model. Arm Holdings plans to launch its own chip this year after securing Meta Platforms as one of its first customers, marking a major shift from its traditional model of licensing blueprints to other companies. This shows how business models can evolve. If TSMC were to do something similar, I would reevaluate my position and might exit.

I was very excited about the new Grok 3 announcement and was contemplating upgrading my X subscription from Premium to Premium+ so I could use it. While I was considering it, X increased the subscription price from around SGD 300 per year to SGD 530 per year after the Grok 3 release. Later, when I logged into the Grok website, I realized that I had access to Grok 3 without upgrading. Perhaps it’s a glitch, or maybe Elon gave me a free trial for a day or two, I have no idea.

This post is to share some tests of Grok 3 and see how good it is. Initially, I wanted to post about Satya Nadella’s announcement regarding the new quantum chip by attaching the tweets. But when I sent the prompt, I forgot to attach the tweet, and it still worked! It looks like Grok 3 is really smart. You can read the prompt and the results below. I’ll also include a comparison with Open AI’s o3 mini.

Satya Nadella’s post:

https://x.com/satyanadella/status/1892242895094313420

Grok 3 prompt:

referring to this tweet. Is it really new state of matter? or Satya is confused? explain in detail

Grok 3 response, with “Think“ turn on:

To address your query about whether Satya Nadella's tweet claiming Microsoft has created a "new state of matter" is accurate or if he’s confused, let’s dive into the details of the announcement and the science behind it. The tweet likely refers to Microsoft’s development of the Majorana 1 quantum chip, which uses a type of material called a topological superconductor. Nadella describes this as an "entirely new state of matter" to enable a leap in quantum computing. Below, I’ll break this down step-by-step to evaluate the claim and clarify what’s really going on.

---

What Does "State of Matter" Mean?

First, let’s establish what a "state of matter" is. In basic terms, a state of matter describes how particles (atoms, molecules, etc.) are organized and behave under specific conditions. The classical states you’re probably familiar with are:

• Solid: Particles are tightly packed and vibrate in fixed positions.

• Liquid: Particles are close together but can move past each other.

• Gas: Particles are far apart and move freely.

• Plasma: A high-energy state where atoms are ionized into charged particles (e.g., in stars or lightning).

Beyond these, scientists have identified exotic states, especially in the quantum realm, such as:

• Bose-Einstein Condensate (BEC): At near absolute zero, certain particles merge into a single quantum state, behaving as one entity.

• Superconductors: Materials that conduct electricity with zero resistance at very low temperatures.

• Quark-Gluon Plasma: A state where quarks and gluons, normally bound in protons and neutrons, roam freely (seen in extreme conditions like the early universe).

These exotic states aren’t new fundamental categories like solid or liquid but are specific arrangements or behaviors of matter under unique conditions, often studied in condensed matter physics.

---

What Did Microsoft Do?

Satya Nadella’s tweet highlights Microsoft’s quantum computing breakthrough with the Majorana 1 chip. This chip relies on topological superconductors, a special type of material that Microsoft claims to have engineered. Here’s what that means:

• Superconductors Recap: Normally, superconductors are materials that, when cooled to very low temperatures, allow electricity to flow without resistance due to paired electrons (called Cooper pairs).

• Topological Superconductors: These are a subset of superconductors with a twist—literally. Their properties depend on topology, a mathematical concept about shapes and connectivity that doesn’t change under stretching or bending (think of a donut vs. a sphere). This topology allows them to host Majorana particles (or Majorana fermions), which are unique because they are their own antiparticles.

• Microsoft’s Breakthrough: They’ve reportedly created these topological superconductors "atom by atom" using topoconductors (likely a catchy term for topological materials). These materials enable Majorana particles, which could make quantum bits (qubits) more stable—crucial for practical quantum computing.

So, the claim is tied to this material’s role in advancing quantum technology. But does it qualify as an "entirely new state of matter"?

---

Is This Really a New State of Matter?

To answer this, let’s evaluate the claim against the definition of a state of matter:

Traditional Definition: A new state of matter would imply a fundamentally distinct phase, like discovering something beyond solid, liquid, gas, or plasma. Topological superconductors don’t fit this mold—they’re a type of superconductor, which is already a known quantum state.

Quantum Context: In condensed matter physics, scientists often describe specific quantum arrangements as "states of matter." For example:

• Regular superconductors are one state.

• Topological superconductors are another, distinguished by their topology and ability to host Majorana particles.

These are more like subcategories within the broader "superconducting state" rather than entirely new phases.

Novelty: Topological superconductors don’t occur naturally and require precise engineering (e.g., extreme cooling and atomic-level control). This artificial creation is groundbreaking, but it builds on existing knowledge of superconductivity and topology, not a radical redefinition of matter’s states.

So, technically, calling it an "entirely new state of matter" stretches the term beyond its classical meaning. It’s more accurate to say Microsoft has developed a new type of material exhibiting a specific quantum state—one that’s rare and valuable but not a fifth pillar alongside solid or gas.

---

Is Satya Confused, or Is This Marketing?

Nadella isn’t a physicist, so it’s unlikely he’s confused about the science in a technical sense. Instead, the phrasing seems deliberate:

• Marketing Angle: Tech companies often use bold, accessible language to hype innovations. "New state of matter" sounds more exciting than "novel topological superconductor," especially for a general audience or investors. It emphasizes the leap forward in quantum computing, which relies on this material’s unique properties.

• Scientific Basis: There’s truth behind it—topological superconductors are a hot topic in physics, and Microsoft’s ability to harness them is a real achievement. Some physicists might even casually call it a "new state" in the context of quantum materials, though not in the classical sense.

However, skepticism exists. Posts on platforms like X suggest:

• One user said, “It’s not a ‘new’ state of matter, it’s a new way to put matter together.”

• Another called it “marketing hype.”

This reflects a divide: the science is impressive, but the language may overreach for effect.

---

Why This Matters

Even if it’s not a new fundamental state, the achievement is significant:

• Quantum Computing: Majorana particles could make qubits less error-prone, a major hurdle in building reliable quantum computers. This could unlock solutions to problems classical computers can’t handle, like simulating complex molecules for drug discovery.

• Materials Science: Engineering topological superconductors pushes the boundaries of what we can create in a lab, adding to our understanding of quantum physics.

---

Conclusion

Satya Nadella’s claim that Microsoft created an "entirely new state of matter" isn’t strictly accurate if we define states of matter by their classical categories (solid, liquid, etc.). Topological superconductors are a specialized quantum state within the known realm of superconductivity, not a groundbreaking new phase. However, he’s not confused—he’s likely using exaggerated language to highlight a real, impressive breakthrough in quantum computing and materials science. The Majorana 1 chip represents a leap forward, even if it’s more about engineering a unique material than discovering a fifth state of matter. So, it’s a bit of marketing flair atop a solid scientific advance—exciting, but not quite as revolutionary as the tweet suggests.

I felt the response was really great. Not only is Grok 3 able to clarify technical terms, but it is also capable of considering Satya's original intention when making the post and viewing it as marketing material.

Here is the response from Open AI's o3 mini:

Satya Nadella’s tweet isn’t claiming that we’ve discovered a completely unknown, naturally occurring state of matter in the same way we think of solids, liquids, or plasmas. Instead, what Microsoft has done is engineer a novel material—what they call a “topoconductor”—that exhibits quantum properties specially tailored for quantum computing.

What’s the Claim?

According to Nadella, after nearly 20 years of research, Microsoft’s team has created a material whose electronic properties don’t fit neatly into the traditional categories (solid, liquid, or gas). Instead, this engineered state is defined by its topological features. In this context, “topological” means that the material’s properties are governed by global features of its electronic structure that protect quantum information against local disturbances. This protection is crucial for building stable, error-resistant qubits.

Is It a New State of Matter?

Here’s where the debate comes in:

• From an Engineering Perspective:
  Microsoft’s achievement is groundbreaking because it demonstrates a new phase or form of matter with properties that are exceptionally well-suited for quantum computing. In that sense, it is “new”—it doesn’t naturally occur in everyday materials and had to be deliberately designed at the nanoscale.

• From a Fundamental Physics Perspective:
  Traditional “states of matter” are defined by the macroscopic properties of systems (e.g., crystalline order, fluidity, or magnetism). Many exotic phases have been discovered over the past century (like Bose–Einstein condensates, superconductors, and superfluids). While the topoconductor exhibits unusual behavior—thanks to engineered topological superconductivity—it might be more accurate to call it a novel quantum phase or engineered material rather than an entirely new fundamental state of matter.

So, Is Satya Confused?

Not really. The language of “new state of matter” is partly a reflection of the excitement and marketing around the breakthrough. Many experts acknowledge that while this material represents a significant leap in quantum materials engineering, it’s a tailored solution aimed at overcoming one of the biggest challenges in quantum computing—error-prone qubits. In other words, it’s less about rewriting the textbook on matter in a broad, universal sense, and more about creating a material with unique, advantageous properties for a very specific application.

In Summary

• What’s New: An engineered material (topoconductor) with topologically protected electronic properties that can support more robust qubits.

• The Claim: It enables a “fundamental leap in computing” by paving the way toward a scalable, million-qubit processor.

• The Debate: Whether it qualifies as a “new state of matter” in the fundamental sense is debated; many see it as an important new quantum phase engineered for a specific purpose rather than a completely new, naturally occurring state.

Thus, Satya Nadella isn’t “confused”—he’s highlighting a major materials science and quantum computing breakthrough, even if the phrasing stretches the conventional definitions of states of matter.

Okay, o3 mini is also able to look up the post without me providing it, so it’s smart too. o3 took about a second to provide the response, but the response was much shorter. I actually prefer Grok 3's response over o3 mini's response.

The Fed minutes are available at this link:

https://www.federalreserve.gov/monetarypolicy/files/fomcminutes20241218.pdf

Here is ChatGPT's response:

Alright, let’s take a closer look at the FOMC minutes and start breaking things down systematically. First, what’s the essence of this document? It’s essentially a record of the Federal Reserve’s deliberations and decisions about monetary policy at their December 2024 meeting. But this isn’t just about their decisions—it’s also about the reasoning behind those decisions, the state of the economy, and their outlook.

First Observation: The Federal Funds Rate Cut

They reduced the federal funds rate by 25 basis points. That’s small, but significant. Why? Well, it shows they’re still in an easing cycle, but the pace is slower. That raises a key question: why slow the pace now? Could it be that they see inflation easing but not fully under control? Or is it a recognition that the economy has shown resilience, and they don’t want to over-stimulate? They explicitly mention the “finely balanced” nature of the decision, so this wasn’t unanimous in spirit, even if most voted for it.

This leads to a deeper inquiry. Inflation has eased, yes, but “remains somewhat elevated.” They seem cautious—almost reluctant—about calling it “tamed.” Why? One possibility is that underlying factors (like wage growth, supply chain disruptions, or geopolitical risks) still pose inflationary threats. But they’re also seeing progress, which could explain why they didn’t opt to hold rates steady.

Second Observation: Inflation Risks and Dynamics

Here’s where things get tricky. They noted progress on inflation but also highlighted risks tilted to the upside. What might these risks be? They mentioned trade and immigration policy changes, which suggests they’re concerned about potential supply shocks or labor market disruptions. There’s also a hint of worry about financial markets—“positive sentiment” and “momentum in economic activity” could keep prices elevated.

Let’s pause here. Is this signaling doubt about the Fed’s ability to fully control inflation with monetary tools alone? It could be. If inflation pressures are structural or tied to external factors, like global supply chains or geopolitical dynamics, their tools might have limits. This raises a philosophical question: should the Fed be more aggressive in such a scenario, or more patient? They seem to be choosing patience for now, likely because the risks aren’t fully realized yet.

Third Observation: Labor Market

The labor market shows signs of “easing,” but the unemployment rate is still low. This duality is interesting. How can we reconcile these two ideas? One explanation is that job openings and quits are declining (signs of cooling), but employment levels remain stable. Another possibility is that productivity gains or changes in labor force participation are masking deeper shifts. Either way, the Fed seems comfortable with the current state of employment, viewing it as consistent with maximum employment goals.

Here, I find myself doubting whether the labor market is really as stable as it seems. Could there be underlying fragility? For example, delinquencies on small business loans are rising, and low- and moderate-income households are feeling financial strains. If those pressures worsen, might they ripple through the labor market? That’s something to keep in mind.

Fourth Observation: Economic Growth and Resilience

Real GDP growth was strong in Q3 and Q4, driven by consumer spending and business investment. This seems like a bright spot. But again, there’s caution. They noted that consumer spending is supported by real wage gains and household net worth, but financial strains on lower-income households could dampen it. Is this a potential tipping point? If lower-income households pull back on spending, might it cascade through the broader economy? It’s possible.

Here’s another thought: they noted that favorable supply developments (like productivity gains) are supporting growth. That’s good, but how sustainable is it? Are these temporary effects, or can they persist? The minutes don’t provide a clear answer, but it’s worth contemplating.

Fifth Observation: International Context

This is a big one. They noted that other central banks are also easing, which creates a divergence in interest rates. This has strengthened the U.S. dollar. But is that a problem? A strong dollar can hurt exports and corporate profits, but it also keeps import prices low, which could help with inflation. The net effect depends on how these forces balance out. I find myself questioning whether the Fed sees this as a short-term issue or a longer-term structural challenge.

Sixth Observation: Balance Sheet and Policy Outlook

The Fed plans to continue reducing its balance sheet, but they expect runoff to end around June 2025. Why June? It’s not entirely clear, but it could be tied to their inflation outlook or liquidity conditions. They’re also signaling a slower pace of policy easing going forward. Is this because they’re closer to neutral? Or is it about preserving flexibility in case inflation resurges? Probably both.

Synthesis

When I step back and try to piece all of this together, a pattern emerges: the Fed is walking a tightrope. They see progress on inflation but aren’t declaring victory. They see strength in the economy but worry about potential cracks. They’re easing policy but want to avoid overdoing it. In other words, they’re hedging their bets.

This makes me wonder: is this approach too cautious? If inflation risks are tilted to the upside, might they be better off holding rates steady? Or is the gradual approach the right call, given the uncertainty? I lean toward thinking their caution is justified—there’s enough uncertainty to warrant a measured approach.

Final Answer

The FOMC minutes reveal a Fed that is cautiously optimistic but deeply mindful of risks. Key takeaways include:

1. A 25-basis-point rate cut signals continued easing but at a slower pace.

2. Inflation progress is evident, but risks (trade, immigration, geopolitical factors) remain.

3. The labor market shows gradual cooling but is still solid overall.

4. Economic growth is strong, but financial strains on lower-income households could pose risks.

5. International developments (rate cuts abroad, a strong dollar) add complexity to policy decisions.

6. The Fed is nearing a neutral stance but retains flexibility to adjust based on evolving conditions.

The overarching theme is one of balance and caution, as the Fed navigates competing priorities in a highly uncertain environment.

Kelvin Learns Investing has finally posted a video on MicroStrategy. In the video, he weighs the pros and cons of investing in MicroStrategy and concludes that he will not be investing in the company, as it is too risky for him.

In the comment section, there are two comments that directly respond to me, which I find amusing. I don’t receive notifications for posts made on other channels, so perhaps the commenters believed that Kelvin would see their comments and share them with me—which, as it turns out, he did.

This post is meant to document my thoughts after reading those two comments.

I think many people don't fully understand my view on MSTR, and honestly, I’m a bit too lazy to explain it in depth most of the time. But here’s the gist:

First off, I’m not certain whether the price of MSTR will crash. In fact, I believe there’s some possibility that its price could go up significantly. However, despite this, I’m still negative about it.

Why? It’s because of the fundamental nature of how MSTR operates. It feels like a scheme where the latecomers end up subsidizing the earlier participants. This creates a cycle: you attract people who don’t fully understand the risks or mechanisms (what I’d call the “stupid money”), and as the cycle goes on, some of those individuals might get smarter and profit—only for more uninformed participants to enter. It’s a never-ending loop of new entrants funding the previous ones, and I’m simply not a fan of this kind of structure.

That said, I might be underestimating just how many noobies there are in this world. These are people who jump in now, maybe make a lot of money, and think they’re clever for doing so. From their perspective, they might assume I don’t get MSTR. But while they might profit, I doubt they truly understand why they succeeded—whether it was skill, luck, or the very nature of the scheme itself.

For comparison, I feel that Bitcoin shares some of these characteristics. There’s an element of subsidization across different generations of holders. However, Bitcoin has been around longer, has weathered its share of volatility, and doesn’t feel as exploitative as MSTR. So I’m relatively more okay with Bitcoin than I am with MSTR.

Lastly, I get that others might have a different philosophy. Some people don’t care whether something is exploitative as long as it makes them money. That’s fair. It’s just not my personal approach, and I don’t expect everyone to agree with my view.

Over the past week, I found myself spending way too much time analyzing MicroStrategy (MSTR). The allure of its "infinite money glitch" was fascinating, but I’ve realized that diving too deeply into this one company isn’t healthy. It can attract unnecessary negativity, and there are plenty of MSTR bulls out there who strongly believe in its narrative. Some even attempted to organize a group attack on my YouTube channel (whatever that means). While I remain bearish on the stock, I genuinely hope their investments perform well over the long term.

Now, shifting gears back to ASML, I came across a post by Mr. WEB that provides a fantastic primer on ASML for those curious about semiconductor engineering and fabrication. I want to focus on one section in particular:

On the revenues side, just like ASML, Coca-Cola sales are predictable. The semiconductor space cannot live without ASML machines, likewise consumers will drink Coca-Cola products everyday.

However on the cost side, Coca-Cola does not need to reinvest much in itself. The syrup sold today is the same one 100 years ago, they don’t even need to increase marketing to attract more demand. Coca-Cola just had to “be smart once”.

In contrast, ASML needs to “be smart everyday”, and that’s a difficult thing to do. The business model is forced to require increasing capital to keep afloat, from the early days of semiconductors this is acceptable because the improvements are huge relative to the level of investments. But now as we are approaching physical limits, it takes more R&D dollars and human resources to have the same percentage improvement.

I agree with the idea that ASML’s business is inherently cyclical—it’s a semiconductor company, after all. The cyclicality can partly be attributed to the "bullwhip effect," a supply chain phenomenon where small fluctuations in end-user demand ripple into larger swings upstream. ASML, as a supplier of EUV and DUV machines, serves foundries like TSMC, Samsung, Intel, and SK Hynix, which in turn supply chips to fabless companies like Nvidia, AMD, Qualcomm, and even big tech firms like Apple and Tesla. Variations in demand for PCs, smartphones, data centers, or Tesla EVs can cascade through the supply chain, causing significant fluctuations for suppliers like ASML.

Does ASML Really Need to “Be Smart Every Day”?

ASML has built its dominance in the lithography space with EUV machines—an engineering marvel. But does it need to keep reinventing itself to remain successful over the next two decades? That’s the question I’ve been mulling over.

In the semiconductor world, Moore’s Law is slowing down. During ASML’s 2024 Q2 earnings call, an analyst pointed out ongoing industry discussions about whether lithography can continue delivering the same level of scaling as in the past 20 years. Semiconductor scaling refers to shrinking transistors to make integrated circuits smaller, faster, and more efficient. This has traditionally driven improvements in speed, power efficiency, and manufacturing costs.

ASML’s CEO, Christophe Fouquet, acknowledged that transistor shrinking has been slowing down for years. However, he emphasized that the demand for greater transistor density remains as strong as ever. Shrinking transistors still represents the best cost-saving method for achieving higher density, and as shrinking becomes harder, its value only increases. Fouquet also noted that as shrinking slows, customers need higher volumes—whether by processing more wafers or stacking more layers on each wafer. Both scenarios drive greater demand for lithography tools, supporting ASML’s long-term growth.

Does Innovation Define ASML’s Future Success?

From a layman’s perspective, I don’t believe ASML needs a groundbreaking invention to sustain its growth. While it continues investing in next-generation tools like High NA and Hyper-NA EUV machines, its current-generation EUV tools already address many of its customers’ needs. Even if ASML halted innovation, customers would likely purchase more of the existing tools to meet demand through increased wafer volume. While this might raise global electricity consumption, it wouldn’t impact ASML’s revenue much.

ASML’s R&D efforts focus not only on improving resolution but also on increasing productivity. Newer EUV machines are larger, more capable, and more expensive, but they deliver a higher return on capital for fabs by enabling faster chip production. This justifies the high price tags, as customers see these machines as valuable investments. By enhancing fab productivity, ASML captures a portion of that value, ensuring customers remain willing to pay for its innovations.

Final Thoughts

ASML doesn’t need to “be smart every day” in the sense of constant reinvention. It has already solidified its dominance in lithography, and with no close competitors, its position in the semiconductor sector remains unchallenged. Yes, the industry is cyclical, but ASML’s strong market position and the value it creates for its customers ensure its financial performance remains robust. Even as transistor shrinking slows, ASML’s ability to drive productivity and deliver value will sustain its growth for years to come.