Field Notes: AI Infrastructure & the New Cognitive Stack

Executive Summary: What I Learned In The Field About Building Infrastructure For AI-Led Intelligence 

The global economy appears to be currently undergoing a structural change that rivals the mechanization of the 19th century and the electrification of the 20th. We are transitioning to an economy defined by electrons generating intelligence. This is not a standard software cycle; it is a capital-intensive industrial supercycle re-platforming the technology stack.

As we usher in 2026, these field notes are a reflection on a number of conversations over the past several weeks with stakeholders ranging from public to private companies and investors. Much of this change will continue well into 2026. 

The analysis that follows dissects four critical pillars of this transformation:

• The Macro-Thesis: Why the “bubble” narrative is a misdiagnosis of a secular supply-demand imbalance, and how “Sovereign AI” and enterprise ROI are creating a floor for demand.
• The Silicon Wars: Strong whispers of having optionality beyond NVIDIA, bolstered by AMD’s chiplet architectures accompanied by “all-in-on-AI” ambitions by CEO Dr. Lisa Su, as well as Google’s vertical integration (TPUs) will/may introduce genuine competition and price deflation to the compute market.
• The Capital Stack: The notion of a $1 trillion financing market is daunting, moving from balance-sheet equity to sophisticated asset-backed securitization (ABS), delayed-draw credit facilities and project finance, amongst other evolving financing mechanisms.
• The Orchestration Layer: The enduring role of software not as a “road-kill” of AI, but as the essential “nervous system” required to aggregate the data that fuels intelligence.

Arvind brings nearly two decades of global operating and investing experience across the U.S., SEA, and South Asia, with a track record that includes four IPOs, six M&As, 30+ growth investments, nine board roles, and more than $19B in outcomes. 

I. The Macro-Thesis 

The debate over an “AI bubble” often conflates financial speculation with industrial reality. Skeptics point to CapEx forecasts approaching $1 trillion by 2030 as evidence of overbuild. However, a demand-supply inspection reveals a different picture.

• Demand Ahead of Supply: Unlike the dot-com fiber glut, which was defined by supply awaiting demand, the current cycle is defined by demand outpacing supply. The “Physics” of building/delivering data centers is being curtailed by supply factors (i.e. the bottlenecks – power availability, CoWoS packaging, transformers and labor). Inability to build fast enough is also restricting the ability to deploy capital and prevents a speculative glut, assuming a robust real-world AI demand. Other drivers of demand include: 
  • Sovereign AI: A unique driver in this cycle is “Sovereign AI.” Nations are funding domestic AI infrastructure as a national security imperative, creating a floor for demand independent of commercial SaaS revenue.
  • Enterprise ROI: The demand signal is transitioning from experimentation to verified ROI. 2025 data indicates that enterprises are moving from pilot programs to full-scale deployments with experienced ROI, see here and here.  Of course, in our own portfolio companies such as Hightouch, Glean, Writer, Sardine, customers are seeing rapid ROI, if weeks if not days. 

Current Thoughts: We are not in a bubble, we are in a state of capital rotation out of legacy on-premise IT and general-purpose CPU compute into accelerated GPU/ASIC compute. We are not seeing a collapse of AI compute demand, we are seeing depreciation and obsolescence risk around capex assets. Below are three observations about the possible risks – 

• The Financing-Useful Life Mismatch – Data centers are long-term project finance/capex deals (20–30 years) to house massive GPU clusters. These clusters’ traditional depreciation schedule pegs GPUs’ useful life to be only 3–5 years, creating added strain on capex operators to find long-duration/-contractual AI demand pipeline/contracts, to profit above and beyond the project finance/capex costs.   However, I will counter the perceived short GPU “useful life” by saying that anecdotally, customers are willing to pay for older H100s at 50% pricing and use it to run inference workloads, thus increasing “useful life” of older chips. This has positive practical implications on cost-to-serve/-to-finance capex.   
• Deflationary Compute – As innovation accelerates, illustrated by NVIDIA’s rapid shift from Hopper to Blackwell to Rubin within two years—the cost of compute is falling sharply. 
• The “Hollow” Middle – This deflationary compute trend threatens mid-tier operators most: without the scale of hyperscalers or the specialization of neo-clouds, they face steep losses if hardware financed at peak prices becomes uncompetitive or if inference pricing collapses with more efficient ASICs.

II. The Silicon Wars: The Era of Specialization

While NVIDIA remains the dominant incumbent, the semiconductor landscape is fractioning, in response to the varied AI workload demand. The “one size fits all” GPU era is ending, yielding to a split between flexible training clusters and cost-optimized inference engines.

The Incumbent: NVIDIA’s Full-Stack Defense

NVIDIA’s dominance is built on the dual moats of CUDA (software lock-in) and NVLink (interconnect performance). Their strategy is “Full-Stack Programmability,” arguing that in a world of rapidly changing model architectures, specialized chips (ASICs) carry too much obsolescence risk. To complement NVIDIA GPUs’ compute strength, in December 2025, NVIDIA also made a move to license Groq’s chip technology, to bridge itself into inference workloads.

However, NVIDIA’s 80% gross margins act as a massive pricing umbrella. Customers are desperate for a “second source” to gain leverage in negotiations and supply chain security.

The “More Than Good Enough” Sniper: AMD

AMD’s Dr Lisa Su is all-in on AI. AMD is the only other vendor which has a comprehensive semiconductor capability (CPUs, GPUs, FPGAs, ASICs) with an integrated systems approach, meeting customers where they are in their workload expansion. AMD is on its 5th generation chiplet technology, giving it a structural advantage as it combines higher yield and lower cost per chip with modular scalability and strong performance-per-watt. 

All of this comes into focus as the workloads start shifting from pre-training to inference, where chip memory becomes critical. It is in that context that AMD’s MI450 could shine, offering 19.6 TB/s of bandwidth (double that of the previous generation) and 432GB HBM4 per GPU—a whopping 50% higher than NVIDIA’s Vera Rubin, creating a ~10TB memory lead at rack scale.  

And AMD’s latest ROCm software stack (ROCm 7), which had received low marks in past, now delivers 3.5x faster inference and 3x faster training than ROCm 6, with expanded support for developer-focused training frameworks and libraries, such as PyTorch and JAX, respectively. Industry observers that I spoke with reported ROCm lags CUDA by 20-25% on certain software benchmarks, and that lag does not seem insurmountable. 

The Vertical Threat: Google TPUs & ASICs

The most potent long-term threat comes from Cloud Service Providers (CSPs) designing custom Application-Specific Integrated Circuits (ASICs), such as Google’s Tensor Processing Units (TPUs). At the recent 2025 AWS re:Invent, Amazon made some announcements as well. 

• • TPU v7 “Ironwood”: Google’s 2025 roadmap features the “Ironwood” TPU, designed for massive scale with optical circuit switching and 1.77 PB of shared memory across pods. This solves the “sharding” problem—the difficulty of splitting a giant model across multiple chips. With shared memory, the entire model can sit in one addressable space. 
  • The Economic Wedge: ASICs can offer 30-50% better power efficiency for specific workloads compared to general-purpose GPUs. In a power-constrained environment, this efficiency advantage translates directly to margin. As the market shifts from training to being inference-heavy, the TCO (Total Cost of Ownership) advantage of ASICs could become difficult to ignore. 

• ASICs Tradeoffs – Programmability: GPUs are not only highly parallel but also highly programmable. This allows developers to innovate at an extreme pace. In contrast, TPU is an ASIC and given the pace of AI innovation, locking in early with a rigid ASIC architecture could be risky. 

Current Thoughts: 

• In consultation with public market semiconductors/related ecosystem analysts, the world is likely going to gravitate towards 60-40; i.e. 60% NVIDIA reliance in near-term and 40% up for grabs for the rest of the semicon chip ecosystem, with AMD and Google TPUs well positioned to gain market share. 
• NVIDIA’s advantages – NVLink remains a key differentiator for multi-GPU LLM serving, providing materially higher bandwidth and lower latency, as KV caches grow with longer contexts and MoE routing 
• AMD’s Competitive Position – MI350-series claims 40% higher tokens-per-dollar vs. NVIDIA B200 and will get better with upcoming MI450 series. Practitioners view AMD as the only alternative credible GPU supplier for inference, with ROCm day-zero support for standard tools and active open-source contributions helping to reduce migration barriers for LLM serving. 
• Google’s Ironwood (TPU v7) is purpose-built for inference workloads such as LLMs, MoE and advanced reasoning. While early, it is being viewed as competitive to NVIDIA on current generation chips, on performance-per-dollar. 

III. Financing the Buildout: The Capital Stack Evolution

Financing a $1 trillion infrastructure overhaul exceeds the capacity of corporate balance sheets. We are witnessing the maturation of AI Infrastructure Financing as a distinct asset class, growing from bespoke loans to standardized instruments often seen in other asset classes.

Project Finance 2.0

The market has shifted toward project finance structures adapted from the infrastructure sector.

• Take-or-Pay Contracts: Deals are underpinned by long-term (3-5 year) “take-or-pay” contracts with investment-grade hyperscalers. These contracts effectively transfer utilization risk from the infrastructure builder to the deployment layer and guarantee revenue/cash flow.
• Forward Flow & SPVs: Developers are utilizing Special Purpose Vehicles (SPVs) and forward flow agreements to fund hardware acquisition. This isolates the asset risk and allows for off-balance-sheet financing, crucial for maintaining corporate credit ratings during massive CapEx cycles.

Securitization (ABS)

The cost of capital is being lowered through Asset-Backed Securities (ABS). 2025 has seen a surge in data center ABS issuance, as operators bundle lease cash flows into rated securities. This signals the market’s transition from “venture” risk to “infrastructure” yield, attracting pension funds and insurers who seek stable returns from the digital economy’s physical backbone. 

The strategic implication: ABS allows for a lower cost of capital (often 100-200 basis points cheaper than private credit) and longer tenor. As the “Neo-Clouds” mature, it is expected they will pivot from high-yield private credit to investment-grade ABS, effectively lowering the cost of compute for the entire industry. 

Current Thoughts: 

• One specific conversation sums it up, “We’re underwriting execution risk, not AI speculation.” Lenders are not betting on whether the AI model works; they are betting on whether the borrower can plug in the servers and keep them running. The credit risk is effectively transferred to the counterparty (Microsoft, Meta, Amazon or pick any large AI-forward large-cap) via the contract. 
• However, the Vendor Financing Flywheel (whereby vendors are financing purchases of their own equipment and booking revenues) introduces circular, interlocking risk, linking equity, debt, and hardware sales (i.e. revenue) in a potentially fragile feedback loop. 
• Geographically, capital terms now hinge on whether financing backs Tier 1 inference hubs with urban latency advantages or Tier 2/3 training sites in rural, power-rich regions—cementing geography as a new dimension of AI credit underwriting.

IV. Software Is Not Dead: It Has Become the Orchestration or Context Layer

The narrative that AI will commoditize software is flawed. In reality, software is evolving into the Orchestration or Context Layer—the essential nervous system required to manage a “System of Action.”

Data Gravity and Governance

AI models are only as good as the data they ingest. Enterprise data remains fragmented and siloed. Software platforms that provide data unification, governance, and lineage are becoming more valuable, not less. “Data gravity” ensures that the platforms holding the transaction records (ERPs, CRMs, companies like our portfolio Databricks, which has a CDW offering) retain the power to orchestrate the AI agents acting upon that data. 

From Chatbots to Agentic Workflows

The industry is pivoting from passive chatbots to active agents. This shift requires sophisticated orchestration to manage complex, multi-step workflows.

• The Governance Moat: Enterprises require deterministic guardrails for probabilistic models. Software provides the logic layer that prevents hallucinations from triggering unauthorized financial transactions or supply chain orders.
• Orchestration-the missing link in Enterprise AI: Orchestration platforms manage the lifecycle of an AI workflow whilst bridging the gap between legacy systems and new AI models, automating the “last mile” of intelligence. They handle prompt chaining, context window management, model routing (sending simple queries to cheap models and complex ones to smart models), and output validation.

Current Thoughts: 

• A comment from a consequential CEO conversation sums it up, “Without software, you have no data; without data, no AI.” Enterprise IT infrastructure is often poor —a fragmented mess of siloed databases, legacy ERPs, and on-premise servers. 
• Putting aside “walled-garden” tactics, the SaaS companies of the last decade are paving the way for AI-native companies to have access to structured data which is a necessary ingredient to the recipe of disaggregating these SaaS companies! 
• However, the generative AI boom is enabling large SaaS vendors to exploit their “data holder” position against AI-native upstarts orchestrating multiple SaaS vendor data. 
• The large SaaS vendors in some cases, are transforming into “Systems of Action,” where the AI provides the probabilistic intent (“I think we should order more steel”), but their software provides the deterministic execution (“Trigger Purchase Order #9923 via SAP API”). 
• This new paradigm of AI-enabled incumbents  is going to be hard to beat, but we are on the lookout for pure-play breakout AI-native companies. This new breed of companies will have some/all of the following control points 
  • Proprietary, compounding data; 
  • System-of-record/-of-action position; 
  • Domain-tuned models to solve “superhuman tasks”; 
  • Robust integrations; 
  • Closed-loop system driving performance

Conclusion

The throughline across these four pillars—macro demand, silicon specialization, financial innovation, and orchestration software—is that the world is quietly constructing the industrial and software base for intelligence itself.

It is also worthwhile highlighting that I currently believe the “AI bubble” rhetoric ignores the fundamental re-platforming of the global economy. We are witnessing a capital rotation where finance is chasing a physical bottleneck (i.e. build out of data centers), driven by AI demand. The winners of this cycle will not be just those with the best models, but those who can master the physical infrastructure (power, silicon) and the orchestration software required to turn electrons into trusted, actionable intelligence. 

Excited to see how 2026 unfolds.