2,655 trillion Korean won.

Approximately 11.68 trillion RMB (~$165 billion). That's Samsung's formally announced ten-year investment total on June 29, with 2,030 trillion KRW going to semiconductor clusters in Yongin and Pyeongtaek, focusing on semiconductor manufacturing, AI compute data centers, and physical AI.

The same day, Beijing-based Hygon Information Technology rang the bell on the Hong Kong Stock Exchange. It opened up 74.56% at HK$199, with a market cap of approximately HK$10.2 billion (~$1.3 billion) and net proceeds of approximately HK$1.415 billion. An AI chip company getting this kind of red-carpet treatment in Hong Kong? Last time that happened was 2024.

Also in late June, OpenAI's custom inference chip Jalapeno (in partnership with Broadcom) was confirmed to have completed tape-out -- going from project initiation to tape-out in just 9 months, with mass deployment planned for end of 2026. Google, Apple, Meta, Microsoft, Amazon, SpaceX -- the list of companies building custom chips keeps growing in TechCrunch reporting.

Three stories dominating headlines on the same day is no coincidence. In summer 2026, the global compute arms race has reached a white-hot stage.

Everyone's Building Chips: Nvidia's "Friends" and "Enemies"

Samsung's investment plan is the largest single commitment in the global semiconductor space today. What does 11.68 trillion RMB look like? Roughly 0.9% of China's full-year 2025 GDP, and nearly 4x Nvidia's FY2025 revenue. That money is going to one goal: narrowing the gap with TSMC in AI chip manufacturing while consolidating leadership in critical components like HBM (High Bandwidth Memory).

Wall Street has already started looking for "the next Nvidia." TechCrunch reported that Micron is being chased by capital -- HBM is a critical component in AI GPUs, and Micron is one of the world's top three HBM suppliers. A Goldman Sachs report estimated Micron and Nvidia would together contribute over 40% of S&P 500 EPS growth in Q2. From GPU sellers to memory sellers, every link in the AI compute supply chain is being revalued by capital.

But what should make Nvidia most nervous isn't Samsung or Micron -- it's its customers collectively "rebelling." OpenAI's Jalapeno is a clear signal: when AI companies' compute spending reaches billions per year, the ROI of custom chips becomes very attractive. Jalapeno is positioned as an inference chip -- inference is currently Nvidia's highest-margin, fastest-growing segment. If OpenAI deploys Jalapeno at scale to replace Nvidia GPUs in its own products, the impact on Nvidia will be direct.

Domestic Compute: The Critical Leap from "Usable" to "Good"

Hygon's IPO is an interesting sample for observing domestic compute progress.

Its prospectus reveals a contradictory reality: on one hand, AI chip demand is exploding and the company's revenue is growing rapidly; on the other hand, the shovel business is getting less profitable -- massive R&D investment, price war pressure, high customer concentration. Domestic compute chip companies are walking a tightrope between "technology catch-up" and "commercial sustainability."

But the trend is clear. DeepSeek-V4-Pro has already completed full-parameter post-training on a domestic compute cluster (Shenzhen Hetao Academy/ShenZhiCheng Compute Network) -- the first third-party institution globally to do so. China's advancing 2.1 trillion RMB national AI data center grid plan explicitly targets 80% domestic chips. When demand, policy, and capital are all aligned, domestic chips going from "usable" to "good" is only a matter of time.

Electricity Is Scarcer Than Chips: The Ultimate Compute Bottleneck

An even more fundamental bottleneck than chips is electricity.

A deep-dive SemiAnalysis report projects that by 2028, installed capacity at US behind-the-meter datacenters will exceed 40GW. Context: a large nuclear power plant has roughly 1GW capacity; 40GW is equivalent to 40 nuclear plants. Grid expansion is nowhere near keeping pace with data center construction.

China faces similar tightness. Multiple regions are already seeing tight data center power quotas, which is one reason "East Data West Computing" and space compute (satellite data centers) are being put on the agenda. When compute is the oil of the AI era, electricity is the oil of the compute era.

The arms race shows no signs of slowing. But history tells us the endgame of arms races tends not to be total victory for one side, but a new equilibrium. In this new landscape, Nvidia won't collapse, but its "dominant" era is ending; domestic chips won't overtake overnight, but their market share will steadily rise; electricity and cooling may become harder bottlenecks to crack than chip design.

Whoever can run the numbers survives the compute wars.