Expert technology analyst Vijay Raina brings a specialized perspective to the intersection of enterprise SaaS and evolving AI architecture. As a leader in software design and infrastructure, he monitors the pulse of the digital economy, specifically focusing on how established giants and emerging enablers are capturing value during this period of intense compute scarcity and architectural transition.
Current compute shortages are forcing many companies to throttle their AI products well below actual capacity. How are specialized manufacturers addressing backlogs in clean room capacity and memory, and what specific steps should firms take to secure advanced logic manufacturing nodes amidst such high demand?
The reality on the ground is that demand for training and inference is so high that many companies simply cannot let their products run at full speed. To address this, specialized equipment makers like ASML and Lam Research are seeing their backlogs inflect as they rush to expand the physical footprint of semiconductor manufacturing. For firms looking to secure advanced logic nodes, the strategy is increasingly about locking in long-term agreements with TSMC, which currently faces no real competition in the high-end space. We are seeing a “buyer of last resort” mentality where companies like NVIDIA are aggressively preempting capacity to ensure they aren’t left behind. Success right now isn’t just about design; it’s about the physical ability to manufacture in ultra-clean environments and securing the massive memory buffers required for LLMs.
While many focus on standalone apps, billions of people interact with AI primarily through integrated search engines. How does this “hidden” AI usage influence advertising conversion rates compared to traditional keywords, and what metrics indicate that a bundled ecosystem provides a more sustainable advantage than raw model performance?
Most people don’t realize that Google Search is already the world’s most widely used AI product, reaching billions of users who may have never even heard of a standalone chatbot. This integration is proving to be a goldmine because AI-enhanced search delivers much higher conversion rates for advertisers than traditional keyword matching ever could. When we look at the sustainability of this advantage, we track user engagement and the depth of ecosystem integration rather than just raw benchmarks. A bundled ecosystem like Google’s or Apple’s wins because it moves beyond a simple “chat box” and becomes a comprehensive assistant that anticipates needs. This creates a moat where the convenience of having your calendar, email, and search connected outweighs the marginal benefit of a slightly smarter standalone model.
Software is shifting from deterministic rules to flexible, contextual interfaces, which threatens traditional seat-based licensing models. Which operational adjustments are necessary for legacy SaaS companies to survive this transition, and why are data infrastructure enablers currently better positioned to capture value than vertical application providers?
Legacy SaaS providers are facing a structural crisis: if AI makes a worker 50% more productive, an enterprise may eventually decide they only need half as many licenses, which destroys the “seat-based” revenue model. To survive, these companies must shift toward value-based or consumption-based pricing, similar to what we see with infrastructure enablers. We currently favor enablers like Snowflake, Datadog, and Palantir because they sit at the foundational layer where data is organized and secured. These firms capture value regardless of which specific app a customer uses, acting as the essential plumbing for the AI era. They are avoiding the “eye of the storm” by focusing on the safety and integration of AI rather than just trying to sell another interface to a shrinking pool of human users.
As AI agents gain access to sensitive personal data like medical records and financial history, how does “trust” alter the barrier to entry for new startups? What specific product features or integrations will eventually drive up switching costs for users who are currently testing multiple platforms?
Trust is becoming the ultimate barrier to entry, as users are being asked to hand over their most intimate data—medical records, payment info, and private schedules—to these agents. This gives a massive head start to established players who already have a track record of security, making it incredibly difficult for a brand-new startup to gain that level of deep access. As users integrate these agents into their personal and professional workflows, switching costs will rise, mirroring the “sticky” nature of the Apple ecosystem. We expect specific features like cross-platform automation and deep integration with existing productivity suites to become the “hooks” that prevent users from jumping to a competitor. Once an agent knows your entire history and manages your life, the friction of “teaching” a new model from scratch becomes a powerful deterrent.
Subscription models alone often struggle to reach profitability in AI, leading to a renewed focus on advertising and transaction fees. How will enterprises balance the productivity gains of AI with the potential loss of per-user license revenue, and what strategies can they use to manage rising inference costs?
The math of a $20-a-month subscription often fails when you consider the massive compute costs of every single prompt, so we are seeing a shift toward advertising and transaction-based revenue. Enterprises are currently walking a tightrope: they want the productivity gains, but they have to manage the fact that running these models is expensive. To manage rising inference costs, many firms are adopting “appropriately-scaled” models—using smaller, cheaper models for simple tasks and saving the expensive, high-powered models for complex reasoning. By absorbing these costs while delivering clear productivity value, software companies hope to maintain their margins, even if the total number of paid human seats begins to moderate over time.
Building out AI infrastructure requires immense capital, often necessitating partnerships between cloud providers and credit markets. How are these projects being structured to provide predictable revenue streams similar to data center REITs, and what benchmarks should investors use to evaluate the execution of these capital-intensive firms?
We are seeing a fascinating shift where AI infrastructure is being financed much like traditional real estate or utilities, involving partnerships between cloud giants, infrastructure specialists, and credit markets. These projects are often structured around discounted cash flow models with highly predictable revenue streams, very similar to how data center REITs operate. For investors, the key benchmarks are execution speed on the product roadmap and actual revenue growth relative to the capital spent. We monitor capital flow into the space closely, but so far, the appetite from strategic partners remains robust because the addressable market for these services is so vast. Large-cap players with strong balance sheets, like Alphabet, are particularly well-positioned because they don’t face the same financing constraints as smaller, speculative firms.
Training models often grabs headlines, but background inference for tasks like security and automated analytics may represent a larger long-term market. How is the expansion into video and virtual modalities shifting demand for compute, and what logic should businesses use when choosing between frontier and scaled models?
While training gets the glory, the real long-term market is in “background inference”—AI running silently in the background to handle security, data analytics, and automated workflows. The move into video and virtual environments is exponentially increasing the demand for compute because these modalities are far more data-intensive than text. Businesses must use a pragmatic logic: use “frontier” models for high-stakes decision-making and creative problem-solving, but pivot to “scaled” or smaller models for repetitive, high-volume background tasks. This tiered approach is the only way to make the economics work as AI moves from a visible assistant to an invisible, omnipresent layer of the enterprise.
What is your forecast for the AI sector?
I believe we are entering a phase of “pragmatic proliferation” where the initial hype around chatbots gives way to deep, integrated utility across every layer of the economy. While the infrastructure players will continue to reap rewards from the compute shortage in the near term, the long-term winners will be the “trust-anchored” ecosystems and the data enablers that solve the safety and integration puzzles. We will likely see a move away from pure network effects toward high switching costs based on personalization and privacy. Investors should watch for the shift from training-heavy demand to inference-heavy demand as AI moves into video and virtual environments, as this will sustain the need for advanced semiconductors for years to come. Ultimately, the companies that can balance high inference costs with clever monetization—like advertising and transaction fees—will be the ones that achieve sustainable profitability.
