Editor's Note: In early 2024, AI was still in a phase of both fervor and uncertainty. At that time, Daniel Gross raised 18 questions on a single page: Where will value flow? Will energy become a bottleneck? Will software engineers be replaced? How will the competitive landscape between nations change?

Looking back two years later, these questions themselves are more enlightening than any specific predictions. AI profits have indeed concentrated at the infrastructure layer—Nvidia emerged as the biggest winner; energy and power quickly became new strategic bottlenecks; API costs plummeted, while computing power, capital, and geopolitical risks continued to amplify.

This article revisits the key questions Gross raised at the time and examines them against the reality of the past two years. It is not only a review of AI investment logic but also a roadmap for observing how the technological revolution is reshaping market structures, industry chains, and the global power landscape.

Below is the original text:

In January 2024, Daniel Gross, then CEO of Safe Superintelligence and now Head of AI at Meta, published an article titled "AGI Trades."

The article was just one page long, listing a series of questions about the potential impacts of AI progress. Looking back over two years later, these questions appear remarkably prescient, even though no definitive conclusions were drawn for each question at the time. Below, we review each of the 18 questions he raised.

Markets

In a post-AGI world, where will value flow?

Currently, value is indeed concentrated at the infrastructure layer—chips, packaging, power, and other areas. Nvidia has captured over 100% of the profits from the AI boom, as many companies are still operating at a loss. This is also clearly reflected in market capitalization changes: Nvidia's market cap increased by $3.2 trillion, rising from $1.2 trillion to $4.4 trillion; in comparison, the gains for cloud platforms have been much more modest (Microsoft up 4%, Amazon up 30%).

In the private market, the valuation growth of OpenAI, Anthropic, and xAI has also been astonishing, but their combined total value increase of $1.4 trillion is still lower than the market cap added by Nvidia during the same period.

This was a critical question at the very beginning of 2024.

What will happen to Nvidia and Microsoft?

Nvidia has performed exceptionally strongly. Its revenue grew from $60.9 billion in fiscal year 2024 to $215.9 billion in fiscal year 2026, nearly tripling.

Microsoft, however, has been less dominant. Azure's growth did accelerate to a 40% year-over-year rate, but from January 2024 to March 2026, Microsoft's stock price rose by only 4%. The market has questioned its annual AI capital expenditure of over $80 billion—when will these investments translate into returns remains unclear.

In this AI "gold rush" of "selling shovels and picks," Nvidia is clearly the biggest winner, while Microsoft's bets on infrastructure have yet to deliver obvious returns to shareholders.

Is copper mispriced?

It was severely undervalued. In January 2024, the price of copper was $3.75 per pound; two years later, it reached a record high of $6.61 per pound.

AI's demand for copper is enormous. For example:

Nvidia's GB200 NVL72 server rack uses over 5,000 copper wires

If fully straightened, the total length exceeds 2 miles

A 100MW data center requires approximately 3,000 tons of copper

Overall, data centers may consume 500,000 tons of copper annually. Some have thus called "copper the new oil." Of course, many other things have also been called "the new oil," as AI infrastructure construction is extremely complex, with bottlenecks at almost every stage. So this analogy should be viewed with caution.

Real Estate

If AI can write all software, will San Francisco become the new Detroit?

It depends on what is meant by "the new Detroit."

AI actually saved San Francisco from becoming a declining city like Detroit. San Francisco is still thriving:

· Office vacancy rates dropped from 36.9% to 33.5%

· OpenAI occupies 1 million square feet of office space

· Anthropic owns a 25-story office building

· Sierra signed a lease for 300,000 square feet of office space

In the first half of 2025, 78% of U.S. AI venture funding flowed to the Bay Area

Of course, there is another side: overall employment in San Francisco remains below pre-pandemic levels, but housing prices remain strong. Thus, it is far from a "shell city." The urban environment has also become cleaner.

How will AI affect wealth inequality?

It is still too early to draw conclusions, as data changes are not yet significant, but some studies are worth noting.

A 2025 IMF study suggested: AI may reduce wage inequality (by automating high-income jobs) but could exacerbate wealth inequality (as gains concentrate among tech company owners).

An OECD study found: Low-skilled jobs saw the fastest wage growth (assemblers +11.6%), while high-skilled jobs saw the slowest (CEOs +2.7%). However, this may reflect minimum wage policies more than AI itself.

In capital markets, concentration is also increasing: The "Magnificent Seven" (Mag7) account for about 32% of the S&P 500's market cap and contributed about 42% of total returns in 2025; meanwhile, massive funding rounds for AI startups (OpenAI $110 billion, Anthropic $30 billion) have also created enormous private wealth for a small number of founders and investors.

Energy & Data Centers

If AI becomes an energy competition, how should one invest?

This assessment is completely correct. AI has indeed become an energy game.

Those who caught this trade made significant profits. For example:

Vistra: +321%, the second-largest gain in the S&P in 2024 (after Palantir)

Constellation Energy: Stock price tripled since ChatGPT's release

NRG Energy: Rose about 95% in 2025 alone

Oklo: Up over 700% in 12 months

Nuclear energy saw a boom:

· Microsoft signed a $16 billion, 20-year PPA to restart the Three Mile Island nuclear plant

· Google signed a 500MW small modular reactor (SMR) agreement with Kairos Power

· Meta signed 6.6GW power contracts with multiple nuclear energy companies

Energy has become one of the most successful investment themes of the AI era.

In the entire data center supply chain, which parts are the hardest to scale 10x?

The bottleneck in the chip industry is CoWoS packaging technology (TSMC's Chip-on-Wafer-on-Substrate).

In the data center field, the biggest bottleneck may be power transformers.

· Delivery lead times approach 3 years

· A 30% supply gap emerged in 2025

· Costs have risen 150% since 2020

This 100-year-old technology has become a key constraint on the speed at which data centers can connect to the grid.

Is coal undervalued?

To some extent, yes, but far less so than copper. Coal prices actually fell about 22% in 2025, with some recovery by early 2026.

Coal companies performed decently:

· Peabody Energy: +34%

· CONSOL Energy: +37%

Meanwhile, U.S. coal-fired power generation increased by 13% by September 2025.

This was particularly evident in states with rapid data center growth:

· Ohio: +23%

· Oklahoma: +58%

Nations

Who are the winners and losers?

The clear winner is the United States.

2024 U.S. private AI investment: $109 billion (China only $9.3 billion)

Cumulative investment since 2013: $470 billion, more than the rest of the world combined

40 significant AI models released by the U.S. in 2024, compared to 15 from China

The game is not over, but for now, the U.S. is the center of AI competition.

What will happen to India's $250 billion GDP export dependency on GPT-4 tokens?

Signs are beginning to show, but it's still early. Hiring in India's IT outsourcing industry has noticeably declined. Between 2024–2025, large IT companies cut about 58,000 jobs, whereas the industry had added 360,000 employees between 2021–2023.

Will software engineers be replaced like typists in history?

Software engineers haven't moved to blue-collar jobs yet, but the career structure is already diverging:

Demand for AI engineers grew 143%

Entry-level hiring at large tech companies fell 25%

Internship positions decreased by 30%

The future choice might be: either move up to become "managers of AI agents" or shift to fields like manufacturing—after all, many factories need people who understand software to automate production lines.

Will there be a large-scale employment program similar to the "New Deal"?

Not yet.

In July 2025, the Trump administration launched the "American AI Action Plan," including:

An AI education executive order

Skills training programs

A $84 million Labor Department apprenticeship grant

But U.S. workforce training spending is only 0.1% of GDP, among the lowest in OECD countries. No plan has yet reached the scale of the original WPA (which employed 8.5 million people).

Is lifelong learning worth investing in?

This is a very abstract and personal question. But my answer is: Yes.

Inflation

If AI is truly deflationary, how would we see the signals first?

The best indicator might be AI API prices.

GPT-4 level inference cost:

Late 2022: $20 per million tokens

December 2025: $0.40 per million tokens

A 50x decrease in three years. This pace even exceeds the decline in PC computing costs or internet bandwidth costs. This is likely to be a leading indicator of service price deflation.

If demand for knowledge products keeps growing while production costs fall, how should we understand deflation?

Although AI API prices have plummeted, AI company incomes are soaring. Price decrease → usage explosion → total spending increases. Meanwhile, SaaS companies are adding 20%–37% "AI taxes" at renewal. Thus, even as the cost of producing software approaches zero, SaaS revenues are still growing.

This is similar to the computing industry in the Moore's Law era: individual products become cheaper, but the overall market size expands.

Geopolitics

Is interconnect really important?

Extremely important.

In large GPU clusters:

30%–50% of training time is spent on communication between GPUs

not on computation

For example:

Google TPUv7 Ironwood uses a 3D torus topology to connect 9,216 chips

Nvidia NVL72 connects 72 GPUs

Thus, interconnect networks are crucial for AI scaling.

If a country has more energy, can it achieve AGI with落后制程 (落后制程 - older process technology)?

Currently, it seems unlikely.

All leading AI chips use 4nm or 3nm processes:

Nvidia Blackwell

Google TPUv7

AWS Trainium3

China's Huawei Ascend 910C (SMIC 7nm) is competitive in inference but requires more chips and more energy for training. Simply increasing energy consumption to bridge the technology gap will eventually run into economic cost limitations.

What is the most likely "Taiwan event"?

The most likely is a blockade of the Taiwan Strait.

Tensions are escalating:

2024: China conducted "Joint Sword-2024B" exercises

2025: "Righteous Mission 2025" deployed over 100 aircraft and 13 warships

27 rockets launched from Fujian, with 10 falling into Taiwan's contiguous zone

Meanwhile, China's 2026–2030 five-year plan began separating the表述 of "peaceful unification" and "unification."

TSMC is also planning ahead: Eight fabs are under construction in Arizona, potentially handling 30% of advanced chip capacity in the future.

But the entire system remains in an extremely fragile balance.