2026 and beyond

Talking about the future is always fascinating, especially when looking at 2025, a year in which AI development accelerated rapidly, no longer prioritizing security concerns tied to effective altruism, but instead emphasizing an innovation war between China and the US. However, few of us realize that in the end, everything is determined by the foundations that build AI itself: energy and hardware components.

I am very confident that China will win this AI innovation race because it has far more energy resources than the US (to the point that the US has had to take over oil reserves in Venezuela and move parts of its oil industry there). This clearly shows that the US is beginning to shift its focus toward energy.

We will focusing around AI, AI progress has been continuous and that what seems like a sudden "explosion" is a continuation of existing trends, leading to a massive research and development (R&D) progress multiplier.

AI progress is primarily bottlenecked by compute rather than the number of researchers. Quantity of researchers, serial speed of thought, computational resources, and AI materials are all inputs to AI R&D, and AI can accelerate all of them.

Disclaimer: The content of this article is part of a thematic research process aimed at identifying patterns, so not everything being written here will necessarily happen in the future.

Now let’s see what I think will happen in 2026 and beyond.

AI Agents are Getting Better

Seeing the development of Claude Code, I am firmly convinced that the evolution of AI Coding Agents will be extremely rapid. Some programmers haven't realized this yet and haven't even tried Claude Code, historically, low alertness usually leads to destruction.

With Claude Code, you only need three basic pillars to survive: Prompts, System Design, and Code Review. I have already seen a phenomenon where people from non-programming backgrounds become "vibe coders" and earn $1M. You can see a similar phenomenon on Shark Tank (https://www.youtube.com/watch?v=t5lN7mRPpw8), which means the AI Ponzi scheme is starting to be felt, beginning from the top.

What Coding Agents are doing now is optimizing by complying with US Executive Order 14110 from the Biden administration to limit FLOPs at 10^26 by creating sub-agents. These sub-agents are currently used by Claude Code to work more optimally on multitasking. However, under the Trump administration, that Executive Order was revoked, resulting in the removal of those FLOPs limits.

The reason the Biden administration implemented FLOPs limits was to prevent rapid developments in bioweapon technology and cyberattacks, but those reasons are being ignored by the Trump administration, which prioritizes US AI Dominance and Innovation.

For this reason, it is highly possible AGI scenario will happen even sooner, where a 4 x 10²⁷ agent becomes a reality. We will see rapid developments not only in bioweapons and cyberattacks, but also in forecasting, healthcare, coding agents, robotics, and even politics:

• Better forecasting model (e.g for Systematic Long-Short trading strategies)
• Advanced Bioweapons, New Palantir innovation
• Better cyberattacks strategies
• Better coding agents for non-programmers
• Better autonomous robotics
• Better law and policy decision making agents
• Better AI Automation, AI Bureaucracy Alpha leading to a unique situation where unemployment may rise
• BS Jobs (bullshit jobs) will be streamlined by AI (e.g Consulting firms, Marketing and Ads agencies)

Advanced Packaging for Semi Conductors and Chiplets

If you are familiar with Moore’s Law, we have traditionally continued to make transistors smaller. In theory, we could make chips larger, but doing so would cause them to no longer fit within lithography machines. This limitation is known as the reticle limit.

If manufacturers produce chips that are larger than what lithography machines can handle, they will encounter a phenomenon known as yield collapse, along with significantly higher production costs. “Yield collapse from bigger chips” refers to a situation in semiconductor manufacturing where the percentage of functional, defect-free chips produced from a single silicon wafer drops dramatically as the physical size (die area) of the chip increases.

The new game is no longer about making a single chip bigger, but about making a chip function as a collection of small chips, known as chiplets, and then integrating them so they behave like one large chip. The process of integrating these chiplets is called advanced packaging.

Competition in advanced packaging has already been underway and is expected to continue between Google and NVIDIA through TPUs, as well as Meta with its own custom chips, all aimed at accelerating AI and machine learning workloads. In this competition, we can observe which players OpenAI aligns with by looking at the widely circulated "OpenAI Money Machine" diagram on the internet. TPUs themselves are Tensor Processing Units, which are specialized chips (ASICs) developed by Google.

In the development of this competition, I highly recommend investing and keeping an eye on Intel and seeing who has the potential to win in this advanced packaging for semi conductors and chiplets battle.

AI Materials

AI Materials phase of the cycle represents a shift from software-defined AI to material-defined AI where the chemical and physical makeup of the hardware is just as important as the architecture itself.

The "material-defined" phase means that innovation is now happening at the atomic level. We are designing new chemical compounds for resins and films just to allow electrons to move fast enough to support 4 x 10²⁷ FLOP agents.

We've optimized the software and the silicon architecture as much as possible. Now, the bottleneck is the "advanced packaging", the way materials like glass, resin, and copper are used to stack chips together.

In the past, AI material companies were treated as commodity chemical vendors. In 2026, they are viewed as strategic sovereign assets.

This is a picks and shovels play in AI, while the world focuses on who has the most GPUs, the real bottlenecks are moving upstream. "moving upstream" refers to a strategic shift in focus from the finished product (like an AI chip or software) back to the fundamental raw materials and chemical processes that make it possible.

To understand "moving upstream" better, it's better to understand the supply chain:

• Downstream (The Mouth): Consumer-facing AI (ChatGPT, Claude Code, AI Agents).
• Midstream (The Middle): Hardware assembly and packaging (TSMC, Nvidia GPUs, Resonac films).
• Upstream (The Source): Raw elements, specialty chemicals, and molecules (Rare earth metals, T-glass fibers, high-purity resins).

For beyond 2026, the "Innovation War" has moved away from who can design the best chip to who can control the very beginning of the supply chain:

• Geopolitical Leverage (Sovereign Assets): If a country controls the Upstream (e.g., China 90% control of Gallium/Germanium refining), they can "turn off the river" for everyone midstream and downstream.
• The "Molecule" is the New IP. You cannot build a 3D-stacked AI chip without a specific Non-Conductive Film (NCF). If a company like Resonac owns the key chemical patents for that film, they effectively own the "gate" to the entire AI industry. Resonac (formerly Hitachi Chemical Co., Ltd.) has a long history in developing and producing NCF materials, which are crucial for high-performance semiconductor chips used in AI applications. The company has made significant investments to increase its production capacity for NCF and thermal interface materials (TIM), which are key components in high-density 2.xD and 3D semiconductor packaging technologies.

This is what makes me believe that China will win the AI ​​race, in terms of material and energy they win against America.

However, it's possible that Japanese companies' patent ownership could make Japan stand out in the midstream of the AI ​​supply chain.

Resonac (A Japan Company) controls approximately 50% of the market for Non-Conductive Film (NCF), which is used to stack memory layers in HBM (High Bandwidth Memory). Without Resonac's film, there would be no HBM3e, and without HBM3e, there would be no Nvidia B200/Blackwell.

For investment advice:

• Downstream: Focus on US-made products
• Midstream: Focus on strategic partnerships with Japanese companies specializing in chemical compounds for AI materials
• Upstream: Focus on companies supplying semiconductor raw materials in China

Natural Gas, Copper, Silver

The transition from software-defined AI to material-defined AI means that the physical constraints of the power grid and hardware components have become the primary bottlenecks for the next generation of AI agents (reaching 4 x 10²⁷ FLOPs):

• Natural gas has shifted from a cyclical commodity traded on weather patterns to a "growth" asset tied directly to AI data center demand.

A major trend for 2026 is the negotiation of fixed-price gas contracts between hyperscalers (like Microsoft or Google) and natural gas producers (like EQT or Comstock). This allows utility-like companies to get a "growth multiple" from investors, as their revenue becomes tied to AI scaling rather than just commodity fluctuations.

• There is an emerging competition for supply between domestic AI data centers and the massive new LNG export terminals in the US.
• Copper demand: The demand isn't just for the chips, but for the massive upgrades required for the power grid and data center cooling systems to support high-density AI clusters. The liquid cooling industry primarily uses all-copper components (e.g copper, nickel-plated copper, and brass) to prevent galvanic corrosion.
• Silver superior conductivity makes it indispensable for the specialized circuitry and advanced packaging required for HBM3e and Blackwell-class GPUs.

References:

• Scott Alexander (astralcodexten)
• James Van Geelen (Founder of Citrini Research)