The agreement will see Meta use tens of millions of Graviton processing cores, according to Amazon Web Services executive Nafea Bshara, who said the contract would run for several years and be worth billions of dollars.

Demand for AI infrastructure is spreading beyond graphics processors made by firms such as Nvidia, while GPUs are essential in training AI models. Companies now need large volumes of central processing units to run trained systems, manage workloads and support AI agents.

Meta said the deal is part of its strategy to avoid relying on one supplier or one type of chip.

“As we scale the infrastructure behind Meta’s AI ambitions, diversifying our compute sources is a strategic imperative,” Santosh Janardhan, head of infrastructure at Meta, said in a statement.

Amazon said Meta chose its latest Graviton5 processor because of its price and performance. The chip is Amazon’s fifth in-house CPU generation and is produced by Taiwan Semiconductor Manufacturing Co.

“We pass that savings on to the customers,” Bshara told Reuters.

He added that most of the chip capacity for Meta would be based in the United States.

The partnership builds on an existing relationship between both companies that dates back several years. Earlier work had focused mainly on cloud services, Amazon’s Bedrock platform and GPU rentals.

For Meta, the latest agreement adds to its list of chip partnerships. The company has already signed major supply deals with Nvidia and AMD, while also working with Arm Holdings.

Amazon, meanwhile, is going deeper into AI infrastructure with both its own silicon and outside partnerships. Earlier this week, it announced another $5 billion investment in Anthropic, which will also use tens of millions of AWS Graviton cores.

Intel said on Thursday that Google will continue using its Xeon processors across a wide range of workloads.

This includes inference and general computing, as companies move from training AI models to running them in real-world applications.

That transition is changing demand. More firms now need chips that can handle steady, heavy workloads rather than short bursts of training. CPUs are becoming more important again, especially for inference tasks and memory-heavy operations.

Google will also adopt Intel’s latest Xeon 6 processors. These chips are designed to improve efficiency and handle larger volumes of data.

According to the companies, they are already being used in Google Cloud’s C4 virtual machines, where they deliver significant cost improvements when running open-source AI models.

At the same time, both firms are working more closely on infrastructure processing units, known as IPUs.

These chips take over tasks such as networking, storage and security, which are usually handled by CPUs. In moving those jobs away, the CPUs can focus on core computing work.

Intel’s chief executive Lip-Bu Tan said: “Scaling AI requires more than accelerators – it requires balanced systems. CPUs and IPUs are central to delivering the performance, efficiency and flexibility modern AI workloads demand.”

The growing use of agent-based AI systems is also pushing demand higher. These systems carry out multi-step tasks and need more background processing power, which usually falls on CPUs rather than specialised accelerators.

For Intel, this is important as the company lost ground earlier in the AI boom to competitors that focused on graphics processing units. Now, it is trying to recover by strengthening its position in general-purpose and infrastructure computing.

The partnership with Google gives it a strong foothold in cloud computing, where demand for AI services is continually increasing.

Intel is also expanding its efforts elsewhere. It recently said it will join a new AI chip project linked to Elon Musk, working alongside SpaceX and Tesla to support robotics and data centre development.

In manufacturing, the company plans to take full control of its Ireland facility by buying back a stake from Apollo Global Management. The site produces Xeon server processors and is paramount to Intel’s supply chain.

Both Google and Intel are expected to highlight their joint work later this month at Google Cloud Next 2026 in Las Vegas, where they will present updates on AI infrastructure, security and edge computing.

That is a 36% year‑on‑year increase from 2025, making this one of the fastest capital expenditure (CapEx) booms in modern corporate history. 

So, let’s discuss. Is this exceptional AI CapEx surge cycle driving productivity in the economy, or are we inflating another technological asset bubble?

The AI CapEx Scale: What’s Happening Now

Across the largest tech firms, the hyperscalers and cloud giants, capital spending is now structural. Amazon, Google, Meta and Microsoft are expected to put hundreds of billions into new infrastructure in 2026, much of it dedicated to specialised computing clusters, advanced networking and data centre capacity. 

The focal point of this spending is not mere servers or office upgrades. It’s data centres built specifically for high‑power compute workloads, facilities optimised for parallel processing at scale. 

These require specialised hardware like GPUs and high‑bandwidth memory, and they draw massive amounts of energy. 

One recent example shows just how strategic these moves have become. Nebius Group signed a multi‑year deal with Meta Platforms worth up to $27 billion to supply dedicated AI computing capacity by 2027, a contract driven by extreme demand and limited supply for high‑performance computing systems. 

Productivity: What the Investment Could Bring

No doubt that enhanced computing capacity enables economic value. Faster processing, more reliable inference workloads, and greater cloud availability can drive:

• Higher labour productivity by automating routine tasks.
• Faster research and development cycles in sectors from healthcare to manufacturing.
• Lower costs for compute‑intensive services, once infrastructure matures and utilisation improves.

For context, the semiconductor industry, a cornerstone of this infrastructure build‑out, is forecast to approach nearly $1 trillion in sales in 2026, with AI‑specific chips maintaining strong annual growth. 

From a macro perspective, such CapEx adds directly to aggregate demand and GDP in the short term. Data centre construction, advanced chip manufacturing, and supporting supply chains all contribute to economic activity that wouldn’t exist without this cycle. 

Bubble Territory: Where the Risks Begin

But there are strong arguments that we are edging into asset inflation rather than productive investment.

First, the pace of spending vastly outstrips current revenue realisation in the economy. Many of these specialised facilities operate at negative operating margins early in their life, requiring ongoing funding before they generate sustainable returns.

Second, a lot of the valuations attached to tech infrastructure assets incorporate lofty future earnings expectations. If those earnings don’t materialise, because adoption slows or competition increases, we could face rapid repricing. 

We’ve already seen some tension in the market, with certain historic investment commitments being scaled back. 

Third, hyperscalers are relying more on external financing even as their own cash flows get tighter. That’s a classic hallmark of an investment boom that may not be fully backed by near‑term productive returns. 

Semiconductors and Data Centres: The New “Oil”?

The analogy of compute as “the new oil” captures two truths:

1. Dependency: Modern AI workloads require massive compute capacity, just as 20th‑century industry relied on petroleum.
2. Infrastructure bottlenecks: Scaling compute, even with unlimited capital, is limited by semiconductor supply, power delivery, and cooling technology.

Already, suppliers like TSMC have posted strong revenue outlooks, showing reliance on advanced chips across the industry.

In parallel, smaller specialist data centre operators, such as CoreWeave, have expanded at a rapid clip. CoreWeave now operates dozens of facilities globally and has become a major supplier for bespoke compute capacity. 

But then, this infrastructure is expensive and energy‑intensive. Many facilities find it hard to break even without long‑term contracts or guaranteed utilisation.

Investment Implications: Winners and Fragilities

From an investment standpoint, certain firms appear ready to benefit if demand holds:

• Nvidia is at the centre of the compute supply chain. Its recent San Jose GTC 2026 forecast shows at least $1 trillion in chip revenue by 2027, driven by demand for next‑generation chips at scale. 
• Other chip designers and foundries stand to gain from backlogged orders and long production lead‑times.
• Data centre REITs and infrastructure funds may see longer‑term cash returns as contracts mature.

On the risk side, overcapacity, falling prices for older hardware, and slower adoption outside of hyperscale use cases are still substantive challenges.

So, Growth Engine or Asset Bubble?

Standing here in March 2026, we see both sides.

On the productivity side, this spending wave is building infrastructure that will underpin major advances in how industries operate. It’s tangible investment in capacity, not just speculation in intangible assets.

On the asset inflation side, the pace and scale of spending go beyond today’s revenue reality. Markets have priced future growth aggressively, which increases the risk of repricing if adoption deviates from expectations.

Now, are we financing a foundation for long‑term productivity, or are we inflating the price of future earnings prematurely?

The company is expected to report a fourth-quarter operating profit of about 16.9 trillion won for the October–December period, according to analyst estimates compiled by LSEG. 

That figure would represent a jump of roughly 160% from a year earlier and place Samsung within touching distance of its 2018 peak, when the memory market last experienced a major price cycle.

What has changed is not just demand, but the structure of it. With manufacturers focusing capacity towards advanced chips for data centres, output of conventional memory has tightened. Prices have responded with unusual force. 

TrendForce data shows DDR5 DRAM prices climbed 314% year-on-year in the fourth quarter, while contract prices for standard DRAM are forecast to rise another 55% to 60% in the first quarter of this year.

That dynamic plays directly into Samsung’s hands. The company is heavily exposed to conventional DRAM, a segment that many rivals had begun to treat as mature. 

“As conventional DRAM prices continue to surge, Samsung – whose production capacity is largely concentrated in this segment – stands to gain relatively more from the current price upcycle,” TrendForce analyst Avril Wu said.

This quarter goes beyond a one-off rebound. Just over a year ago, Samsung’s leadership was apologising publicly for weak performance as it fell behind SK Hynix in supplying high-bandwidth memory to Nvidia. 

Today, the tone is different. On Friday, executive chairman Jun Young-hyun told investors that customers had described Samsung’s next-generation HBM4 chips by saying, “Samsung is back.”

The competitive backdrop explains why that is important. SK Hynix completed what it described as world-first HBM4 development in September 2025, doubling bandwidth and cutting power use by 40%. 

By the end of last year, it had already sold out its entire 2026 supply to Nvidia. Micron, meanwhile, has told investors that tight memory conditions could last beyond 2026, with chief executive Sanjay Mehrotra warning that the company expects to meet only half to two-thirds of demand from several major customers.

At CES 2026, chipmaker Nvidia unveiled its Vera Rubin platform, confirming that the next generation of its systems will rely on HBM4 memory. Nvidia said the Vera Rubin architecture is in full production and on track for launch later this year, underlining how critical reliable HBM supply has become.

Samsung’s expected profit surge shows this bigger change. Some analysts have already lifted their fourth-quarter forecasts above 20 trillion won, betting that price momentum in traditional memory has been underestimated. 

Looking further ahead, market forecasts reveal Samsung’s operating profit could exceed 100 trillion won this year, more than double last year’s level, if pricing remains firm.

Investors have largely embraced the turnaround. Samsung shares rose 125% last year, their strongest annual gain in 26 years, although they dipped 2.1% in early Tuesday trading as the wider market paused after a rally.

Risks have not disappeared. Lee Min-hee of BNK Investment & Securities cautioned that higher chip prices could cool demand for consumer devices and flagged “risks of a demand slowdown” as data centres rely more on debt to fund expansion. 

Samsung itself has acknowledged the limitations on its mobile business, where rising component costs are squeezing margins. “As this situation is unprecedented, no company is immune to its impact,” co-chief executive TM Roh said, adding that the fallout looks “inevitable”.

Even so, a memory market once dismissed as cyclical has become essential to the next phase of global computing. 

In a post on his social media page, Trump stated: “With this TRUTH, I put all Countries with Digital Taxes, Legislation, Rules, or Regulations, on notice that unless these discriminatory actions are removed, I, as President of the United States, will impose substantial additional Tariffs on that Country’s Exports to the U.S.A., and institute Export restrictions on our Highly Protected Technology and Chips.”

Trump argues that DSTs are designed to harm American technology firms, while allowing Chinese competitors to avoid similar treatment. His position revives an old fault line between Washington and its allies. 

During his first term, he had also threatened countries such as Canada and France with tariffs for pursuing similar tax regimes. In February this year, he ordered U.S. trade officials to reopen investigations into countries levying DSTs against U.S. tech giants.

Over 20 nations, including France, Spain, Italy, India, Kenya, and the United Kingdom, have introduced DSTs ranging between 2% and 7.5% of gross revenue from digital advertising, marketplaces, and user data monetisation. These policies primarily affect firms such as Google, Meta, Apple, and Amazon, which dominate the global digital economy.

While proponents argue that DSTs ensure fair taxation of multinational platforms profiting from their markets, the U.S. government sees them as discriminatory. Officials believe they tilt the playing field against American companies while giving an advantage to rivals, particularly those from China.

Beyond DSTs, the United States has grown more wary of the European Union’s landmark digital regulations, the Digital Services Act (DSA) and the Digital Markets Act (DMA). The DSA, enforced in 2024, compels platforms to remove illegal content, boost transparency, and share data with regulators. 

The DMA aims to curb anti-competitive behaviour by major “gatekeepers” such as Google and Apple, forcing them to open up their platforms and reduce self-preferencing.

Washington interprets these moves as non-tariff trade barriers. Reports state that Trump’s team has even considered sanctions on EU officials responsible for enforcing the laws, a step that could further strain the $1.7 trillion transatlantic trade relationship.

Trump’s latest warning goes beyond tariffs as he also threatened to restrict exports of advanced semiconductors and artificial intelligence chips, a measure that could disrupt supply chains worldwide. Companies like Nvidia, which play a central role in AI development, could be caught in the crossfire.

The U.S. and EU conduct more than $4.2 billion in trade daily, and a recent agreement capped U.S. tariffs on most European goods at 15%. Introducing new tariffs or export controls would escalate tensions and risk retaliation from allies.

Efforts to resolve the tax dispute at the multilateral level have also faltered. The OECD has been pushing for a global framework to replace DSTs with a uniform system for taxing multinational profits. However, the U.S. remains resistant, fearing it would lose its own taxation rights under the proposed arrangement.

With both sides unwilling to compromise, the digital tax fight appears set to intensify. Trump’s latest threat raises the prospect of a trade confrontation both with rivals, and with long-standing allies who see DSTs as a matter of fairness in the digital economy.