Silicon Starvation: Amazon Feasts on Legacy GPUs

Key Points

• Hyperscalers like Amazon are shifting away from commercial GPUs toward proprietary chips such as Trainium 3, pushing Taiwanese suppliers to raise shipment volumes sharply.
• Amazon is pricing reserved GPU capacity higher while its AWS custom silicon pipeline reaches an estimated $225 billion in committed customer revenue.
• Taiwan Semiconductor Manufacturing Co. benefits as the sole foundry for major hyperscalers' custom chips, running 3nm capacity at full utilization with expanding margins.

Hyperscalers are quietly ending their dependence on legacy graphics processing unit (GPU) makers by aggressively migrating to building their own chips. For years, the major cloud providers operated essentially as toll collectors. They bought commercial GPUs off the shelf at a premium and rented out that compute power to enterprise clients.

That dynamic is breaking down rapidly. Cloud giants recognize that relying on third-party designers for foundational hardware creates unacceptable margin compression. The physical economy of AI infrastructure is shifting away from off-the-shelf components toward proprietary systems that cloud providers control end-to-end.

Taking the Big Bite: A Volume Spike Disrupts Supply

The clearest evidence is showing up in Taiwan's supply chain, where order volumes for custom chips are spiking well beyond normal levels. This structural pivot directly threatens the market share of legacy chip designers while creating an extraordinary demand supercycle for the pure-play foundries—companies that manufacture chips but don't design their own—executing these manufacturing orders.

Investors navigating the second half of 2026 should recognize that the most lucrative infrastructure investments are no longer the companies designing generic chips. The real capital is flowing toward cloud providers building proprietary chip ecosystems and contract manufacturers who physically print that silicon for them.

The Premium Platter: Amazon Prices Out Competitors

An aggressive move in this transition recently came from Amazon.com, Inc. NASDAQ: AMZN. Amazon Web Services (AWS) recently initiated an urgent supply chain adjustment, instructing Taiwanese server component manufacturers to hike third-quarter 2026 shipment volumes by 20% to 30%. This volume pull-forward targets Amazon's proprietary Trainium 3 infrastructure. Early production for Trainium 3 ramped up in May 2026, and nearly all incoming capacity is fully reserved by core enterprise clients.

Amazon is using its dominance in cloud architecture to steer the market toward its proprietary hardware. It recently executed a 20% price increase for reserved commercial GPU capacity. By raising the cost of third-party compute, Amazon makes those chips financially impractical for large-scale workloads, effectively herding enterprise clients into its own Trainium and Inferentia ecosystem.

This strategy is already bearing fruit at an institutional scale. Under the internal moniker Project Rainier, leading artificial intelligence developer Anthropic is actively scaling its compute architecture on hundreds of thousands of Trainium chips. Anthropic openly states that current architecture planning based on a 10x growth multiplier is insufficient, validating the urgent need for localized inference workloads. This aggressive deployment pushes the AWS custom silicon pipeline to an estimated $225 billion in committed customer revenue.

From a valuation standpoint, Amazon absorbs broader market tech rotations with remarkable resilience. Amazon's share price has maintained a steady uptrend in the $240-$245 range, and is up roughly 10% year-over-year. While e-commerce margin compression remains a subtle macro headwind, the stability of cloud revenue and the margin expansion inherent in owning the hardware stack easily offset retail pressures. Trading at a trailing price-to-earnings ratio of 29x, Amazon commands a premium pricing environment entirely driven by these infrastructure expectations.

The Kitchen's Cut: Printing Silicon at a Premium

Designing custom application-specific integrated circuits (ASICs)—chips designed for one narrow purpose rather than general use—is highly profitable. However, producing them requires specialized fabrication capabilities that cloud providers simply do not possess.

Taiwan Semiconductor Manufacturing Co. NYSE: TSM is the definitive pure-play beneficiary of the hyperscaler pivot. It is the sole supplier of custom chips designed by Amazon, Google, and Meta NASDAQ: META.

Foundries represent the ultimate toll road in the semiconductor market. TSMC is currently running its 3nm fabrication capacity at 100% utilization. Because no other manufacturer can reliably yield chips at this advanced node, the manufacturer wields absolute pricing power. Management recently implemented price hikes of 10% to 15% on its 3nm advanced nodes, expanding gross margins to an exceptional 66.2%.

Wall Street analysts are rapidly recalibrating models to account for this pricing leverage. Bank of America recently raised its price target for TSMC to $590, projecting 2027 capital expenditures of $78 billion. Analysts expect gross margins to sustain above 66% through 2028 as combined wafer capacity scales toward 400,000 units per month. S&P Global also revised its outlook to positive, acknowledging the unassailable moat built around leading-edge node manufacturing.

TSMC is actively mitigating geopolitical supply chain headwinds. The company recently executed a 10-year strategic partnership with Amkor Technology NASDAQ: AMKR to expand advanced chip packaging in Arizona. This initiative cements a localized U.S. infrastructure footprint, pairing efficiently with a recent move to raise the quarterly dividend payout to $1.1136 per share.

Shares currently trade in the $435 to $452 range, advancing on heavy volume. The current 36x trailing price-to-earnings multiple represents a substantial expansion from the five-year median of 23x, reflecting the premium investors are willing to pay for absolute market dominance.

Hungry Investors: Asset Managers Gorge on Foundries

Capital allocation flows highlight strong institutional conviction in the foundry-and-custom-silicon thesis. Tier-one asset managers are heavily accumulating shares of primary manufacturers.

Capital Research Global Investors added 6.77 million shares of TSMC, while Capital World Investors accumulated an additional 5.48 million shares in the most recent quarter. Corporate insiders mirror this confidence, executing 77 purchase transactions over the trailing six months, compared with a single sale.

Bearish conviction against this trade is virtually nonexistent. Amazon's short interest is negligible at 1.01% of the public float. While regulatory filings show planned stock sales by Amazon executives, including a 20,500-share open-market sale by the CEO of Worldwide Amazon Stores, these are routine wealth-management exercises rather than bearish signals. Options chains for both Amazon and TSMC reflect elevated implied volatility, squarely targeting the upcoming mid-July 2026 earnings reports as investors position for forward guidance revisions.

Clearing the Table: Rotating Capital to the Winners

Market data adds necessary nuance to the hardware replacement narrative. Base-layer demand for commercial GPUs remains robust for raw, foundational training workloads. The true shift is materializing in the inference market, which now accounts for two-thirds of all compute spend. As artificial intelligence applications transition from initial training to daily execution, hyperscaler custom silicon is monopolizing localized inference workloads.

Custom integrated circuits carry a 44.6% forward compound annual growth rate, dwarfing the 16.1% growth rate projected for legacy solutions. The data suggests a rapidly evolving hybrid ecosystem where legacy units handle generalized training, while proprietary chips handle the highest-margin, highest-volume inference.

Investors seeking optimal exposure to the next phase of the computing cycle should consider moving capital toward the vertically integrated hyperscalers driving this transition and the sole-source foundries that print the physical architecture.