Anthropic’s $7bn Chip Hunt Tests the Economics of Gulf AI Investment

Saudi Arabia and the UAE are spending heavily to become global suppliers of AI computing power. Anthropic’s push into proprietary silicon raises a harder investment question: can data-centre returns keep pace with increasingly rapid hardware replacement cycles?

Anthropic’s abandoned attempt to acquire AI-chip start-up MatX for about $7 billion carries implications well beyond Silicon Valley. For the Gulf, it exposes an emerging financial risk beneath one of the region’s biggest technology bets: the accelerated economic depreciation of the AI processors and computing hardware powering its new data centres.

Reuters reported that Anthropic considered acquiring MatX, founded by former Google TPU engineers, before shifting towards a possible partnership. MatX is separately seeking funding at a valuation of around $4 billion, while Anthropic is recruiting semiconductor specialists and examining several chip developers as it builds an internal silicon capability.

The motivation is increasingly economic. Anthropic is simultaneously committing enormous sums to external computing capacity, including a reported six-year, $45 billion agreement for around 460MW from Nscale using Nvidia’s Vera Rubin systems.

At that scale, even relatively small improvements in the cost, power consumption or utilisation of each AI workload can justify investment in proprietary processors.

Anthropic is also following a broader industry direction already established by Google, Amazon, Meta and others: leading AI companies increasingly want greater control over the hardware beneath their models.

For Gulf investors, that changes the equation.

The Gulf’s Compute Bet

Saudi Arabia and the UAE are positioning computing capacity as a new strategic infrastructure industry.

Saudi Arabia’s HUMAIN and AMD announced plans to invest up to $10 billion to deploy as much as 500MW of AI capacity over five years. HUMAIN’s Nvidia programme separately targets up to 500MW, beginning with an 18,000-unit GB300 Grace Blackwell system. Both are expansion targets rather than currently operating capacity.

The Saudi strategy is already broadening. HUMAIN, AMD and Cisco subsequently announced a joint venture targeting up to 1GW by 2030, beginning with a planned 100MW deployment, while Qualcomm is targeting 200MW of inference infrastructure from 2026 alongside an AI engineering centre in Riyadh.

Abu Dhabi is pursuing even greater scale. Stargate UAE is being developed as a 1GW AI cluster within the planned 5GW UAE-US AI Campus, with the first 200MW under construction towards planned delivery in 2026.

The opportunity is substantial. Demand remains strong: Nvidia’s latest outlook pointed to continued rapid growth in AI infrastructure spending rather than an imminent collapse in accelerator demand.

The investment risk is subtler.

MEO Analysis: The “Stranded Compute” Risk

A data-centre building, grid connection, fibre network and cooling system can remain productive for years or decades.

The accelerator hardware inside it operates on a very different economic clock.

MEO defines stranded compute not as an obsolete data centre, but as accelerated economic depreciation of installed processors before investors have fully captured the returns originally expected from them.

The distinction is important.

Older GPUs do not suddenly become useless when a new generation arrives. They can move into less demanding training or inference workloads, or remain profitable at lower prices.

But if newer accelerators perform substantially more computation per megawatt or per dollar, the market value and rental economics of older machines can weaken quickly.

That creates an unusual asset-duration mismatch:

long-lived infrastructure surrounding short-lived technology.

The financial transmission is straightforward:

faster hardware depreciation → lower achievable pricing or utilisation → higher replacement capex → weaker free cash flow and project IRR.

For Gulf investors, that is more important than whether one particular processor architecture wins.

MW Measures Capacity. Utilisation Determines Returns.

The AI infrastructure race is frequently described through GPU numbers and megawatts. Neither determines profitability on its own.

A 500MW facility operating close to capacity under multi-year contracts with creditworthy customers can produce attractive infrastructure economics.

The same facility carrying expensive processors without sufficient utilisation can destroy capital.

Anthropic’s Nscale deal illustrates why contract structure matters almost as much as hardware selection. Long-duration customer commitments can transfer part of the technology-cycle risk away from the infrastructure owner by providing greater revenue visibility through several years of hardware depreciation.

Gulf AI projects should therefore increasingly be judged on four variables:

utilisation, contracted revenue, replacement capex and upgrade flexibility.

Investors should ask who pays when processors need replacing, who carries their residual-value risk, whether contracts survive a hardware refresh, and whether power and cooling infrastructure can support future architectures.

This is particularly important as financing moves deeper into AI equipment. Reuters Breakingviews has highlighted efforts to develop hundreds of billions of dollars of financing around AI hardware, including structures that may depend partly on assumptions about the residual value of Nvidia accelerators.

The question is increasingly not whether AI demand exists, but who absorbs depreciation if hardware economics change faster than expected.

Stranded Compute Does Not Mean Stranded Infrastructure

There is also an important counterargument.

Better processors can actually increase the value of high-quality Gulf data centres.

If a future accelerator delivers considerably more compute from the same megawatt of electricity, scarce grid capacity becomes more productive. Revenue per MW can rise even as individual servers are replaced.

That means the underlying powered infrastructure — land, electricity, fibre and cooling — could retain or increase its strategic value while the accelerator layer turns over rapidly.

The Gulf’s comparative advantages therefore remain powerful.

Its challenge is to avoid allowing most technology-cycle risk to sit on regional balance sheets while the highest-margin intellectual property remains elsewhere.

From Compute Capacity to Value Capture

Saudi Arabia is already beginning to address that issue. Qualcomm’s collaboration includes a semiconductor design centre intended to develop local engineering capabilities alongside infrastructure investment.

That may prove strategically important.

Gulf investors do not need to replicate TSMC’s fabrication model to capture more of the AI value chain. More durable opportunities could lie in chip design, specialised software, networking, advanced packaging, model optimisation and equity participation in semiconductor technology, alongside data-centre ownership.

The capital-allocation principle is straightforward: regional investors should seek assets capable of generating returns across multiple generations of processors, rather than relying excessively on the scarcity premium attached to one generation of hardware.

Anthropic’s MatX discussions therefore do not undermine the Gulf’s AI strategy. Even companies developing proprietary processors continue to commit tens of billions of dollars to external compute, demonstrating how large global demand remains.

But the investment standard is becoming tougher.

The next phase of Gulf AI development should be measured less by headline GPU counts and more by how effectively projects recover capital, maintain utilisation and upgrade through successive silicon cycles.

The key question is no longer simply how much compute Saudi Arabia and the UAE can install.

It is whether the region can keep earning attractive returns after today’s chips are replaced.

Gulf Capital Shifts from AI in Riyadh to Energy Deals in Abu Dhabi

From Resort to City: Can Egypt Build Year-Round Tourism Economies?