AI trade is still on?

A great conversation spanning the AI value chain and an indicative view of the trends for the rest of the year. This conversation is from RiskReversal Media podcast - "AI investment Promises Made, Will they be Kept" and the conversation is between host Dan Nathan and Parshv Shah from Kingdom Capital.

Capital Allocation & Macro Investment Thesis

A central theme across the AI value chain today is whether the scale of AI infrastructure investments will ultimately yield sustainable returns. There’s growing concern that the industry may be overcapitalizing at the infrastructure level, particularly in compute. While there's a clear constraint on the availability of advanced compute, especially for large model training, the limiting factor now appears to be the capital required to deploy this compute at scale. This marks the classic Jevons Paradox in action—where increased efficiency in compute only accelerates overall demand. The broader debate on AI's capital intensity, sustainability, and return profile is still evolving, with no clear consensus.

Power constraints and move towards higher throughput per MWs

At the recent NVIDIA GTC 2025 conference, one striking data point was the revelation that ResNet inference is now 100% GPU-based, signaling a complete migration away from CPUs for this class of workloads. This reflects how deeply GPUs have penetrated both training and inference, reinforcing their role as central compute engines in AI infrastructure.

Elsewhere, Alibaba’s latest earnings underlined that the AI bubble is not evenly distributed. Some business units, especially tied to AI monetization, are underperforming, while others continue to scale. This heterogeneity across verticals reinforces the view that AI’s commercial maturity is still highly context-specific.

One of the dominant GTC topics was the networking substrate for data centers. There’s growing emphasis on optical networks and co-packaged optics (CPO) as the next leap in energy-efficient data transmission. While copper interconnects and air-cooled data centers remain the norm, they're beginning to hit thermal and power efficiency limits, especially at hyperscale.

From a product cadence standpoint, all eyes are now on NVIDIA’s Rubin architecture, though we’re still in the deployment phase of Blackwell GPUs. These product launches will define compute availability well into 2026.

Meanwhile, Capex growth across hyperscalers is now in the single-digit range, sparking concern that the pace of infrastructure growth may be slowing. While some see this as a sign of overcapacity building up, others argue we haven’t yet extracted full value from the AI workloads that will justify these builds.

Structural Shifts in Hyperscalers & AI Demand

Despite aggressive investments, hyperscaler stocks—particularly Microsoft Azure, Amazon AWS, and Google Cloud—are under pressure. The market is responding to decelerating cloud revenue growth, juxtaposed with rising Capex commitments, especially for AI workloads.

Interestingly, generative AI hasn’t displaced core search yet. According to the latest earnings call data, Google Search is still growing at around 12% year-on-year, generating close to $300 billion in annualized revenue. This suggests that the economic moat around search remains intact. However, there's an ongoing innovation dilemma: how to introduce generative interfaces while preserving monetizable search intent.

The S&P 500 and NASDAQ 100 are now highly concentrated, with the top stocks making up 33% and 50% of the index weight, respectively. This includes the usual suspects—Apple, Microsoft, NVIDIA, Alphabet, Meta, Amazon, and Tesla. Despite macro concerns, many fund managers believe this concentration is justified by the outstanding cash flows, robust product pipelines, and deep AI moats these companies command. Still, the index structure poses systemic risks in the event of any correction in the AI trade.

Hardware Supply Chain & Custom Silicon

Within the supply chain, Broadcom stands out. The company saw a 40% surge in its stock price in December 2024, largely due to optimism around custom silicon revenues. The expected Total Addressable Market (TAM) for AI accelerators and ASICs is estimated to grow from $60 billion to $90 billion over the next few years. Despite this optimism, recent price corrections reflect investor demand for proof of delivery.

From a financial planning perspective, companies are being impacted by the elongation of depreciation cycles, now stretching to 4–6 years. This affects earnings and introduces delays in Capex reallocation. The earnings compression seen in some quarters could eventually serve as a signal for future Capex planning inflections.

Server & Memory Economics

In the AI server market, Dell and Supermicro are two key players. Dell remains a volume leader, but Supermicro is gaining share, thanks to faster product iterations. That said, margins are under pressure across the board, with Dell’s server margins currently sitting around 22%. Dell also struggles with end-to-end integration; it hasn’t been able to attach storage and networking components effectively to its compute offerings.

On the memory side, Western Digital and Micron continue to lead. Micron, in particular, posted a strong earnings beat, but its NAND flash margins remain under pressure, as manufacturing complexity increases and ASPs compress. The memory and storage sector overall remains less profitable than compute, though critical for performance scalability.

Security & Software Layer

A major downstream effect of AI adoption is the expansion of the enterprise attack surface. As companies deploy generative models internally, from chatbots to code generation tools, security risk vectors have multiplied. There is a pressing need for purpose-built AI security layers, especially for LLM monitoring, prompt injection protection, and model access control. This domain represents a new greenfield opportunity for cybersecurity vendors.

AI SaaS Adoption & the Monetization Puzzle

Salesforce and Adobe represent two of the most active SaaS players incorporating AI. Both have released AI copilots and creative AI tools within their respective suites. However, the key issue remains: revenue uplift hasn’t followed user engagement. This highlights a classic Innovator’s Dilemma—SaaS companies must embed AI to defend their existing product moats, yet monetization strategies remain uncertain. Some tools enhance user experience, but few have proven revenue-generating at scale.