Sands Capital is a staff-owned growth manager founded in 1992 by Frank Sands Sr. in Arlington, Virginia, running high-conviction concentrated portfolios of innovation-led growth businesses with about $46bn in client assets. Its "What We Think" column publishes deep research on technology, healthcare and emerging supply chains.
This report says the recent AI chip stock sell-off is a technical correction, not a fundamental downturn, and the author remains bullish long-term. The author thinks the market overreacted, as Google, Microsoft, and Amazon's cloud businesses are growing fast with strong AI demand. Key holdings: TSMC saw 78% profit growth, boosting 2027/2028 forecasts; Micron signed 16 take-or-pay contracts through 2030, locking in demand; CXMT (a Chinese chipmaker) expansion is misread—its new capacity mainly serves domestic needs, with limited global impact.
One-sentence summary: The author believes the recent sell-off in the AI infrastructure sector is a technical correction rather than a fundamental inflection point, maintaining a long-term bullish stance [Bullish].
The report argues that the recent sell-off in the AI infrastructure sector is primarily driven by technical pressures, crowded positioning, and debates over return on investment, rather than demand deterioration. The author notes that the PHLX Semiconductor Index surged approximately 80% in the second quarter (its strongest quarterly return since the index was created in 1993), before falling over 20% from its June peak by the end of July, with all 30 constituent stocks dropping below their 50-day moving averages. This extreme trajectory made the sector highly sensitive to narrative shifts, potentially amplifying indiscriminate selling. The author states, "The correction followed an extraordinary advance. The PHLX Semiconductor Index gained approximately 80 percent in the second quarter, its strongest quarterly return since the index was created in 1993," meaning: "This correction occurred after an unprecedented rally. The PHLX Semiconductor Index rose about 80% in the second quarter, the strongest quarterly return since the index was established in 1993." Key catalysts include Meta Platforms discussing leasing out excess computing capacity, Moonshot AI releasing the low-cost open-source model Kimi K3, and the IPO of China's ChangXin Memory Technologies (CXMT) , which intensified market concerns over computing capacity oversupply and declining profitability. The author believes the market reaction is disproportionate to the fundamental impact.
Second-quarter results from hyperscale cloud providers show strong demand, with AI infrastructure economics improving. Alphabet, Microsoft, and Amazon reported accelerating cloud revenue growth, with stable or expanding profit margins. Google Cloud revenue grew 82% year-over-year, with EBIT margins expanding 300 basis points quarter-over-quarter, indicating that new capacity is being rapidly absorbed and AI workloads are supporting, rather than undermining, cloud profitability. Amazon disclosed that its custom chip business has reached an annualized revenue of $25 billion, with Trainium chips now supporting over half of the usage on the Amazon Bedrock platform and becoming the primary inference chip for Anthropic. AWS maintains high utilization and expanded margins to 39% by directing growing token demand toward cost-effective self-developed chips. Alphabet's TPU chips also contributed to a strong quarterly performance. The author believes custom chips are becoming a cost advantage and a significant revenue source.
Despite market concerns, AI computing capacity supply remains constrained, with high long-term demand visibility. Alphabet raised its 2026 capital expenditure outlook by $15 billion, citing the ability to accelerate capacity delivery, and repeatedly described the business as supply-constrained, even temporarily accessing SpaceX computing capacity at a premium. Amazon similarly noted that meaningful capacity is already booked through 2028, emphasizing that new capacity is only brought online when customer demand visibility is high. Taiwan Semiconductor Manufacturing Company (TSMC) saw 78% of its earnings growth drive further upward revisions to 2027 and 2028 earnings forecasts, reinforcing strong demand for advanced process semiconductors. Micron Technology has signed 16 strategic customer agreements, most of which are take-or-pay contracts through 2030, while SK hynix has signed five-year agreements with 10 customers. These commitments improve visibility into demand, pricing, and production planning, while phased capacity expansion reduces the risk of speculative oversupply seen in previous semiconductor cycles.
The report believes the current sell-off in the AI infrastructure sector is a technical correction rather than a fundamental inflection point, reaffirming conviction in the long-term opportunity for AI infrastructure. Institutional bias note: Sands Capital is a long-term bullish holder of AI infrastructure positions, and its analysis naturally carries a bias toward defending its holdings. Readers should be aware that the report's judgment of "market overreaction" may underestimate the uncertainty surrounding intensifying competition and returns on capital expenditure.
The report argues that the low cost of the Chinese open-source model Kimi K3 actually proves that compute demand will not decrease, because cheaper costs stimulate broader usage. The author notes that Kimi K3 required 2.8 trillion parameters to achieve competitive performance, demonstrating that "lower cost does not necessarily mean less compute." The report further explains that efficiency gains lower the cost per unit of intelligence but spur greater consumption—"the relevant question is therefore not whether models become cheaper, but whether falling costs expand usage faster than efficiency reduces the compute required for each task." The author concludes that evidence increasingly points to the latter.
The report argues that supply concerns triggered by CXMT's IPO are overstated, as its new capacity is primarily absorbed domestically and does not involve the core high-bandwidth memory (HBM) bottleneck. Key data and logic are as follows:
| Factor | Report's Assessment | Supporting Data |
|---|---|---|
| Global memory supply growth | Low 20% range | CXMT grows much faster than the industry, but the increment is mainly absorbed domestically |
| Global memory demand growth | Over 30% | AI infrastructure and increased memory content in end markets |
| CXMT's impact on global markets | Limited | New capacity likely replaces Chinese imports rather than loosening global supply |
The report emphasizes that the capacity constraint on high-bandwidth memory (HBM) is the key pricing factor, and CXMT lags behind industry leaders by approximately 2-3 years in this area. The author points out that HBM consumes 3-4 times more wafer capacity per bit than standard DRAM, is more complex to manufacture, and has lower yields. As a result, manufacturers prioritize HBM production, squeezing standard DRAM capacity. CXMT's increase in standard DRAM capacity "does little to relieve the constraint currently setting prices." Additionally, new capacity construction has shifted from past brownfield projects to greenfield factories requiring 2-3 years to build, meaning no material relief before 2028. The industry has consolidated to three players—Samsung Electronics, SK Hynix, and Micron Technology—which lock in high-margin opportunities through long-term agreements. CXMT may become a more significant supply source by the end of the decade, but the report views it as "a potential later-cycle risk rather than a challenge to its near- and medium-term foundation."
By refuting the two bearish arguments of "Chinese low-cost models + capacity expansion," the report reinforces confidence in sustained AI infrastructure investment. Readers should note that the author is a long-term bull on AI infrastructure, and the analysis naturally tends to downplay supply-side risks while emphasizing demand-side elasticity—the long-term threat from CXMT is explicitly acknowledged but pushed to the distant future, while the persistence of the HBM bottleneck is used as a core argument.
The author believes that the recent pullback in the AI infrastructure sector has driven valuations to more attractive levels, and the market's skeptical view of the cycle conflicts with fundamental evidence from cloud growth, backlogs, margins, utilization, customer commitments, and persistent capacity constraints. The author's original statement: "The recent retracement across AI infrastructure has driven valuations to more attractive levels and embedded a more skeptical view of the cycle. In our assessment, that skepticism conflicts with the evidence from cloud growth, backlogs, margins, utilization, customer commitments, and persistent capacity constraints." The author acknowledges that the pace of investment and rapid technological evolution will continue to generate volatility, but argues that the divergence between sentiment and technicals versus the fundamental outlook may create opportunities, and maintains confidence in specific providers of compute, memory, networking, and power infrastructure.
The author explicitly states that this is a moment for fundamental investors to exploit the disconnect between price and fundamentals and maintain conviction supported by evidence. It should be noted that this is a self-defense from a position-holder's perspective—Sands Capital itself is heavily weighted in AI infrastructure, and its judgment naturally carries a bias toward maintaining its positions. Readers should independently assess whether the "evidence" it cites (such as cloud growth and capacity constraints) is sufficient to offset concerns about return on investment and excess risk.
| Ticker | Direction | Author's One-Sentence View | Key Data |
|---|---|---|---|
| Alphabet | Hold & Watch | Cloud business growth is accelerating, TPU chips contribute strongly, and AI infrastructure economics are improving | Google Cloud revenue grew 82% YoY, EBIT margin expanded 300 bps QoQ; 2026 capex outlook raised by $15 billion |
| Microsoft | Hold & Watch | Cloud revenue growth is accelerating, AI workloads support profitability | Cloud revenue growth accelerated, margins stable or expanding |
| Amazon | Hold & Watch | Custom chip business annualized revenue reaches $25 billion, Trainium chips support over half of Bedrock platform usage | Cloud revenue growth accelerated, margins expanded to 39%; capacity already booked through 2028 |
| TSMC | Hold & Watch | Strong demand for advanced process semiconductors, earnings growth drives upward forecast revisions | 78% earnings growth drives upward revisions for 2027/2028 earnings forecasts |
| Micron Technology | Hold & Watch | Long-term take-or-pay contracts lock in demand, HBM capacity constraints are key to pricing | Signed 16 take-or-pay contracts through 2030 |
| SK hynix | Hold & Watch | Long-term agreements lock in high-margin opportunities, HBM capacity constraints persist | Signed five-year agreements with 10 customers |
| Samsung Electronics | Hold & Watch | After industry consolidation, three players dominate; HBM capacity constraints are core | Industry consolidated to three players; HBM consumes 3-4x wafer capacity per bit vs. standard DRAM |
| CXMT | Not specified | Short-term threat is overstated; new capacity is mostly absorbed domestically; HBM lags by 2-3 years | Global memory supply growth below 20%, demand growth above 30%; CXMT grows much faster than industry but incremental output is absorbed domestically |
| Meta Platforms | Not specified | Discussion of leasing out excess compute capacity became one of the sell-off catalysts | Raised market concerns about compute oversupply |
| Moonshot AI | Not specified | Released low-cost open-source model Kimi K3, but the author believes low cost does not mean lower compute demand | Kimi K3 required 2.8 trillion parameters to achieve competitive performance |
| Anthropic | Not specified | Amazon's Trainium chip becomes its primary inference chip | No specific data |
| SpaceX | Not specified | Alphabet temporarily accessed its compute capacity at a premium, reflecting supply constraints | No specific data |