Recent discussions have increasingly focused on SMCI, with clients raising deeper questions about whether to maintain positions or hedge against the stock. Clients are asking , “Legal talk would redefine the partnership between SMCI and NVIDIA?“, “$7B of finished goods and inventory—how much of that is truly transferable value versus written down?“ and many more. The current landscape suggests that the concerns surrounding the company have evolved from speculative claims into a more structured credibility challenge.

Source: US Senate

Last year in March 2025, NVIDIA CEO Huang told Bloomberg that there is no evidence of diversions when it comes to its Grace Blackwell chips. Furthermore, he added that Nvidia’s customers are well aware of the restrictions and that they are careful about it when selling to third parties.

Now let’s rewind SMCI’s 2024 events to-date.

The shift in market sentiment for SMCI is defined by these key developments:

• Initial 2024 Allegations: Hindenburg Research first brought forward a series of operational red flags.

• Auditor Resignation: Ernst & Young (EY) subsequently stepped down, stating they could no longer rely on the company’s internal governance or the representations made by management.

• 2026 Legal Developments: A 2026 indictment has introduced criminal export-control allegations involving senior individuals linked to the firm.

While these legal filings do not necessarily validate every specific claim made by short-sellers, they appear to reinforce the market’s focus on governance and compliance. What might have been viewed as isolated noise in 2024 is now being re-evaluated as an early indicator of a more systemic internal control problem.

Investors also need to stop pretending the governance debate is separate from the commercial debate. EY resigned in 2024 saying it could no longer rely on management’s and the Audit Committee’s representations. SMCI’s special committee later said it found no substantial integrity concerns, said tone at the top was appropriate, and said EY’s conclusions were not supported by the facts it reviewed. But BDO later concluded SMCI did not maintain effective internal control over financial reporting as of June 30, 2025, citing multiple material weaknesses, including timely identification and disclosure of new related-party transactions. SMCI also disclosed that it received DOJ and SEC subpoenas in late 2024 following the August 2024 short-seller report.

That does not prove every historical allegation against the company. But it does mean the market is rational to apply a governance discount. And that matters because the key question in AI infrastructure is no longer just “Who can ship?” It is “Who can be trusted to ship, support, document, finance, and defend the shipment?”

For added context, this is not the first time internal-controls issues have been part of the SMCI story. In 2020, the SEC charged Super Micro and its former CFO over widespread accounting violations, and the company agreed to pay a $17.5 million penalty.

Ultimately, the progression from external accusations to an auditor’s exit and, finally, to a criminal indictment has shifted the narrative. The market is no longer just looking at a matter of public perception, but is now weighing whether these events point toward fundamental issues within the company’s leadership and regulatory adherence.

In today’s article, we’ll dig into the questions that have dominated recent client conversations:

1. Following the recent charges against senior-linked employees for illegal GPU exports, the question for every stakeholder is: does this indicate an isolated compliance failure, or a systemic governance issue that could redefine the partnership between SMCI and NVIDIA?

2. Is SMCI a real platform company with durable execution advantages, or was it just the fastest conduit for Nvidia demand?

   1. If Nvidia never formally restricts SMCI but quietly shifts engineering priority, allocation timing, and support toward Dell, HPE, and ODMs, how quickly would that show up in win rates—and would the market recognize it before revenue breaks?

3. If large buyers of Blackwell or Rubin clusters are already rationally shifting 10–20% of future awards away from SMCI—either by splitting vendors or avoiding sole-source exposure—does that reveal that SMCI’s moat was never as deep as bulls believed, but instead dependent on speed, flexibility, and trust that can be re-underwritten elsewhere?

4. Inventory Is Not Just Big at SMCI — It Is Slower, More Finished, and More Exposed Than Bulls Admit: SMCI reportedly carries ~$7B of finished goods and inventory—how much of that is truly transferable value versus written down?

   1. If SMCI were to enter a distressed or restructuring scenario, who are the natural acquirers of its assets—Dell, Hewlett Packard Enterprise, Lenovo, or ODMs like Quanta Computer / Wiwynn—and what would each actually want (customers, design wins, or inventory)?

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Read more

On March 2, 2026, NVIDIA announced a $4B investment split between Coherent and Lumentum. The headline number is actually two coordinated moves — $2B into Coherent and $2B into Lumentum — each tied to multi‑year purchase commitments and capacity/priority access aimed at scaling co‑packaged optics (CPO).

Coherent makes the linkage explicit in its own 8‑K, disclosing that the collaboration expands NVIDIA’s access to “five additional product families related to co‑packaged optics,” while NVIDIA has already telegraphed CPO networking as a path to scaling “AI factories” via Quantum‑X Photonics and Spectrum‑X Photonics.

In real CPO deployments, lasers are often pushed out of the hot ASIC package into external laser source modules that are thermally isolated and field‑replaceable, feeding silicon photonics engines near the switch. The Lumentum leg completes the picture: it looks like NVIDIA underwriting the laser bottleneck that CPO architectures expose.

Lumentum had already begun expanding U.S. manufacturing for CPO‑relevant laser capacity, and NVIDIA’s investment reads less like a sudden pivot than a deliberate sequence: supplier builds the capacity roadmap → NVIDIA finances and reserves it as CPO ramps.

The investor takeaway is blunt: this is not just demand pull, it’s supply gating; CPO shifts what must be secured upstream (lasers, advanced optics, integration capacity rather than just pluggable modules); and the fact that the deals are nonexclusive signals a multi‑sourcing strategy, not vertical capture.

Secondly, let’s look at the data outlined in the McKinsey projections. It illustrates exactly why hyperscalers are being forced to rethink their architectures.

Source: McKinsey Report on Networking Optics

By 2029, total transceiver demand is projected to double to 28 million units, with a staggering 41% concentrated strictly in the 1.6 Tbps segment. While highly engineered OSFP pluggables can technically survive at these 1.6T speeds, they will not thrive.

Pushing legacy form factors to 1.6T exacts a massive power penalty—often requiring 30 to 40 watts per module—which pushes conventional air cooling to its absolute physical limits. If the industry attempts to meet this 11-million-unit demand exclusively with these power-hungry pluggables, data centers will rapidly exhaust their facility power budgets. Therefore, this chart does not just map the growth of the transceiver market; it pinpoints the exact capacity threshold where the massive power tax of 1.6T pluggables forces CPO out of R&D and into mandatory deployment.

We do not have to guess CPO timeline. The public markets are already flashing concrete leading indicators for CPO deployment:

• NVIDIA CPO switching systems commercialization: NVIDIA’s technical blog states customers and commercial availability targets of early 2026 for Quantum-X InfiniBand photonics switches and 2H 2026 for Spectrum-X Ethernet photonics switches

  

  .

• Broadcom CPO shipping (hyperscaler early deployments): The Next Platform reports Broadcom’s second-generation CPO (“Bailly”) Gen 1 began shipping to hyperscaler customers in 2023 and provides concrete per-port power comparisons versus pluggables.

  

• Hyperscaler bandwidth mix shift (demand pull): McKinsey projects hyperscalers will shift ~87% of back-end optics to 800G and above by 2029, with 1.6T transceivers >40% of that back-end demand.

• Supply pressure (forcing function): McKinsey also projects 800G transceiver production could fall 40–60% short of demand through 2027, and 1.6T supply shortfalls of 30–40% through 2029—which tends to accelerate architectural substitutions (LPO, on-board optics, CPO) where feasible.

Our clients are also wondering about Applied Optoelectronics and record braking performance. We will also comment on AAOI’s Moat.

In today’s article, we’ll address client asked questions that determine whether CPO becomes a lasting structural shift. Note: This piece focuses on recent market catalysts rather than foundational engineering. For a breakdown of the technical contrasts between CPO and traditional pluggables, please refer to our 2024 architectural deep dive.

1. If pluggables (OSFP/QSFP-DD) can scale to 1.6T, What is the bull case for Co-Packaged Optics (CPO)?

   1. Applied Optoelectronics’s pluggable Moat:

      1. Is AAOI’s 400mW pump laser a credible pivot—or too small?

      2. Why would a hyperscaler choose AAOI for transceivers over Coherent/Lumentum/Broadcom?

   2. Discuss key CPO adoption concerns and give our take

2. Adoption pace (scale-up vs. scale-out):
   How quickly will co-packaged optics (CPO) be adopted in scale-out data center fabrics versus scale-up accelerator interconnects, and what specific workloads, network tiers, or deployment environments will drive the earliest production use?

3. Value chain and value capture:

   1. As the industry shifts from front-panel pluggable optics to CPO/near-ASIC optical engines, how does the content and profit pool redistribute across the value chain?

   2. What are the second-order effects of CPO on today’s optical and networking supply chain—who gets disintermediated, which components become bottlenecks, and what does this mean for incumbents in transceivers, components, and manufacturing services?

4. Upstream equipment suppliers that benefit:
   Which upstream equipment and “picks-and-shovels” suppliers are likely to see outsized demand from CPO adoption and why?

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Q1. Pluggables vs. bull case for CPO & Our Take on Concerns

At first glance, the existence of 1.6T pluggables raises a fair question: if the interface already scales, why introduce architectural complexity with co‑packaged optics? The answer lies not in whether pluggables work, but in whether they remain the most efficient system-level solution as bandwidth, density, and AI-driven scale accelerate. To evaluate the bull case for CPO, we need to look beyond module capability and examine system constraints.

We need to look at OSFP not as a permanent cure, but as the ultimate, highly engineered band-aid.

OSFP is brilliant, and it is the exact reason data centers can survive 800G and early 1.6T speeds right now. By making the pluggable module slightly larger and strapping a massive heat sink to it, OSFP buys the industry time.

When you see Applied Optoelectronics’ stock rally, it can be confusing at first. If NVIDIA and others are pushing hard toward co-packaged optics, why is a pluggable transceiver supplier like AAOI gaining traction?

The key distinction is:

“Can a pluggable module exist at 1.6T?” → yes.

“Is a pluggable architecture still the best system architecture at 1.6T+ AI scale?” → often not.

We don’t need CPO because pluggables “can’t do 1.6T.” We need CPO because at 1.6T and beyond, pluggables can become too power-hungry, too thermally constrained, and too electrically lossy for the densest AI fabrics.

Applied Optoelectronics’s Moat in Pluggables

Today, On March 9th, Applied Optoelectronics announced that it received its first volume order for 1.6T data center transceivers from a major hyperscale customer possibly suspected Amazon and the stock jumped! Our clients are wondering ,Why would a hyperscaler choose AAOI for transceivers over Coherent / Lumentum / Broadcom? And if their InP manufacturing capacity is a real moat?

The customer order details say,

AAOI occupies a specific position in the data center value chain. They don’t manufacture switches, servers, racks, or cabling infrastructure; instead, they focus on the optical transceivers that move data between those systems.

AAOI’s CFO outlined a mid‑2027 monthly transceiver revenue run-rate target near $378M/month (mix: $91M 100/400G, $217M 800G, $71M 1.6T), based on three hyperscale customers. That’s the opposite of a company being “immediately disrupted” by CPO—AAOI is effectively saying the next 18–24 months are dominated by pluggable ramps.

As we discussed regarding CPO and 1.6T pluggables, the entire industry is facing a component urgency and future shortage. The ultimate chokepoint is not assembling the metal boxes; it is the fabrication capacity for Indium Phosphide (InP) laser diodes.

• Unlike many module vendors that simply buy lasers from third parties and assemble them, AAOI manufactures its own lasers in-house.

• Because they control the raw “fuel” (the lasers) required for both modern pluggables and future CPO external light sources, hyperscalers are flocking to AAOI to bypass the broader supply chain gridlock.

Why hyperscalers pick AAOI

Read more

Last time, we argued that agentic AI is structurally deflationary for SaaS. But that only holds if the agent is consistent, auditable, and able to operate over time without re‑asking the same questions or repeating the same mistakes. That takes us to the real infrastructure question behind the SaaS compression narrative: Memory. As clients start underwriting agents as systems of record for decisions and actions—not just assistants—many are now focused on the next market: the memory/state layer that makes agents dependable, and how large that spend category can become.

In the last 4-6 months, the questions we’re getting from clients have shifted from “can the agentic model answer this?” to “can the agent run the workflow end‑to‑end, reliably, repeatedly, and at a predictable unit cost? What’s infrastructure behind them?”

In practice, that immediately turns into a memory problem: why does the agent re‑ask the same clarifying questions, forget constraints mid‑task, or behave inconsistently across sessions? And once agents move beyond a single response into minutes‑long or even tens‑of‑minutes execution, memory stops being a “nice feature” and becomes the substrate for continuity, auditability, and control.

Recent market data confirms that the value capture in the AI stack is shifting upstream to the component level. As highlighted by financial reporting on the “AI Memory Boom”, memory manufacturers like SK Hynix and Kioxia are seeing valuation surges exceeding 1,000%—outperforming even the logic designers. This “supercycle” is driven by the acute shortage of HBM3e (High Bandwidth Memory), which is now the primary constraint on GPU/TPU production.

While hyperscalers (Google/Meta) and chip designers (NVIDIA) fight for dominance, the memory fabs have secured effective pricing power, with production sold out years in advance, effectively turning HBM into the industry’s most scarce commodity.

Other Data points that explain why memory is suddenly first‑order (not optional):

Tokens = COGS, and re-deriving context is expensive at scale. Anthropic’s pricing guide uses a support-agent example with ~3,700 tokens per conversation and estimates roughly ~$37 per 10,000 tickets (example math depends on model/pricing tier), illustrating how quickly costs compound when the system repeatedly re-loads history instead of remembering key decisions.

Agents are now designed to run for “several minutes” to “tens of minutes,” which forces durable state (memory) and job semantics. OpenAI’s docs explicitly position Background mode for long-running tasks and note that agentic systems can take several minutes, while “Deep research” tasks can take tens of minutes

Source: Open AI

Cloud platforms are explicitly monetizing memory/session primitives, which is a strong signal that the market is treating memory as core infrastructure. Google’s Vertex AI Agent Engine/Agent Builder materials indicate Sessions and Memory Bank (along with Code Execution) moving to metered billing starting Jan 28, 2026 (per pricing/release notes).

Taken together, these developments show memory moving from an implementation detail to a monetized, governed infrastructure layer—the prerequisite for agents that run longer, cross tools and models, and operate inside regulated enterprises.

Per Mordor, The Agentic AI Orchestration And Memory Systems Market size is estimated at USD 6.27 billion in 2025, and is expected to reach USD 28.45 billion by 2030, at a CAGR of 35.32% during the forecast period (2025-2030).

Source: Agentic AI Orchestration And Memory Systems Market

In today’s article, we’ll address these client asked questions related to Memory and Agentic AI Infrastructure:

1. AI agent memory infrastructure: What is AI agent memory infrastructure, and how do its different memory layers change the overall demand for memory across the stack? Why do multiple memory layers are required?

2. Build vs buy (memory/state infra): When should an AI infrastructure company build its own memory/state layer vs adopt a vendor—and what are the hard gates that decide the answer?

   1. Pilot, Growth and Large Scale Scenarios with Bedrock AgentCore Memory pricing vs. Build in-house

   2. Analysis of Menlo and Battery’s reports and takeaway

3. Architecture + tokens (state placement in a token-priced world):

   1. (Token ROI): In a token‑priced world, how should teams budget tokens across instructions, retrieved context, state, and tool outputs—especially for long‑horizon agents where “more context” can increase cost and reduce reliability?

   2. (DeepSeek tie): If model efficiency keeps improving (e.g., DeepSeek‑style cost/perf shifts), what changes in the infra playbook: more on‑prem/on‑device inference, bigger working contexts, different caching strategies—or does “state outside the model” still dominate for correctness?

4. Who’s leading: Which players are leading in agentic infrastructure memory?, and what have recent projects/launches demonstrated about what really works at scale?

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Q1. AI agent Memory Infrastructure: Why Multiple Memory Layers Are Required?

Over the past year, the conversation around agentic AI has shifted from model quality to operational durability. Investors are no longer asking whether agents can generate text—they’re asking whether agents can run workflows, maintain continuity across sessions, survive failures, and operate under governance constraints. That shift brings one infrastructure layer to the center: memory. In agentic systems, memory is not about “remembering a chat”—it is the mechanism that makes behavior consistent, enforceable, and economically scalable.

AI agent memory infrastructure refers to the systems and patterns that enable an AI agent to store, retrieve, manage, and reason over persistent knowledge across interactions, contexts, and tasks. Memory allows agents to behave statefully—remembering facts, preferences, plans, past interactions, world state, and long-term learning outcomes.

For production-grade agents (e.g., in enterprise automation, ITSM, knowledge work agents, robotics), memory infrastructure is a foundational capability that supports:

• Context continuity: retaining user or environment history across sessions

• Personalization: storing user preferences and profiles

• Causal reasoning: chaining past events into current decision-making

• Long-term planning & learning: accumulating experience and improving over time

• Knowledge grounding: integrating domain knowledge, databases, and ontologies

Memory infrastructure is not just a datastore; it’s a multi-layered retrieval and reasoning system.

Let’s look at the diagram below and walk thorugh an example to understand this better.

Source: Agentic Memory Aurimas

To understand above, let’s run this diagram through a typical Agentic AI user request. Let’s say our user request is:

“Schedule ‘Tech Deep Dive’ on Dec 31 for 60 minutes. It must be at least 30 minutes after Marcus’s ‘Code Review Session’ and at least 30 minutes before any event I already have.”

Read more

On January 12, 2026, Anthropic released Claude Cowork as a “research preview”—positioned explicitly as “Claude Code for the rest of your work”—inside the Claude Desktop macOS app. Cowork’s distinctive claim is not “better answers,” but permissioned execution: you point it at a folder, and it can read, create, edit, move, rename, and (critically) delete files while also taking over a browser workflow when allowed.

What made Cowork land as an investor-grade catalyst was the combination of (a) outcome execution in a mainstream desktop form factor and (b) the speed at which it was shipped. Anthropic staff described Cowork as built in roughly “a week and a half” / under two weeks, with reporting that Claude itself coded “pretty much all” of it and that developers were running ~3 to 8 Claude instances in parallel to implement and debug features. In other words, the product was a live demo not only of automation, but of software production velocity.

One mid-tier product marketer reacted,

Cowork was initially gated behind Anthropic’s power-user Max subscription, which is explicitly tiered at $100/month (“Max 5x”) and $200/month (“Max 20x”), advertising 5x and 20x more usage per session than the Pro plan. Within days, it broadened: by January 16, 2026, Engadget reported Cowork was opened to $20/month Pro subscribers as well (with Anthropic cautioning Pro users may hit limits sooner).

The equity market response was immediate and measurable. In the days following Anthropic’s announcements, RBC highlighted a broad software selloff, with shares of Salesforce, Workday, Intuit, and Snowflake falling roughly 6% to 13% as investors re-ran the playbook on where pricing power sits when an agent can do multi-step work across tools. That repricing was not confined to single names: Bloomberg reported that a Morgan Stanley basket of software/SaaS names was down ~15% year-to-date, after declining ~11% in 2025—i.e., a two-year narrative overhang now had a new “proof point” to trade on.

You can see the same risk premium in sector-level numbers. The iShares Expanded Tech-Software Sector ETF (IGV) showed a NAV of ~$98.35 and ~‑6.96% NAV return year-to-date as of January 16, 2026, with ~$6.678B in net assets, ~67.9M shares outstanding, and an indicated portfolio P/E ~42.03 (a reminder that even after drawdowns, software remains a long-duration asset class that is hypersensitive to perceived unit-economics regime shifts).

Two additional quantitative context points matter because they explain why this time the market cared about “agents” rather than treating them as another feature cycle. First, AI is now a line-item budget reality: Investopedia cited a Jefferies survey indicating ~12% of IT budgets dedicated to AI “this year,” up from ~6.5% the prior year, alongside an expectation of ~10% cloud spending growth—meaning buyers are actively reallocating dollars, not just experimenting. Second, capability is still uneven: TIME reported that Opus 4.5 (the model powering Cowork) completed 9 of 240 tasks in the Center for AI Safety’s Remote Labor Index—useful for investors because it frames the near-term impact as pricing/seat pressure and workflow redesign, not instantaneous full autonomy.

So the core underwriting problem is not whether Cowork “works.” The underwriting problem is whether outcome-executing agents change (1) where pricing power migrates, (2) which layer becomes the enterprise control plane, and (3) what observable metrics prove real spend compression rather than narrative-driven volatility.

Clients have been asking whether Claude’s rapid enterprise penetration—sparked by the Cowork release—is a durable inflection or just another short-lived AI hype cycle.

In today’s article, we’ll address client asked questions that determine whether this becomes a lasting structural shift in software economics:

1. What is Agentic AI? Why did late 2025—rather than earlier GenAI cycles—become the point when “agents” started to matter for software valuations, and what system-level change triggered that shift?

   1. Review of McKinsey, BCG & MIT, ISG late 2025 reports

2. When Anthropic released Claude Cowork—an agent that can read/write files, drive browsers, and complete multi-step tasks—the market immediately sold off a wide range of enterprise software stocks; which parts of the software stack actually face durable pricing compression, and which are structurally insulated?

   1. Impact on Salesforce, Adobe, Atlassian, Workday, ServiceNow, Snowflake, Datadog, Twilio, and Intuit

3. What concrete evidence would confirm that agentic AI is reducing some SaaS spend rather than merely increasing usage after the initial Cowork-driven shock?

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Q1. What is Agentic AI? Why did late 2025 become the pivot point?

Agentic AI is best understood as the shift from “AI that answers” to “AI that executes.” Instead of producing text outputs that humans must manually turn into work, agentic systems can hold state, plan multi-step workflows, take actions across tools, and verify results—often producing real artifacts (spreadsheets, decks, tickets, code changes) as the deliverable.

What changed in late 2025 is that this was no longer a lab prototype or a brittle demo: vendors began shipping agent loops (plan → act → verify → iterate) inside permissioned environments with sandboxes, connectors, and governance primitives.

Source: ISG Market Report

Let’s review leading industry reports first before we talk about why late 2025 became a pivot point.

According to MIT and BCG’s Nov 2025 report, organizations with the highest level of agentic AI adoption are far more likely to see increased applications of the technology across various modes of decision-making. The report said, 76% of respondents to survey say they view agentic AI as more like a coworker than a tool.

But the largest impact of agentic adoption is augmenting human judgment: 79% of extensive agentic AI adopters say they are investing in using AI to generate insights for a human decision maker.

Source: MIT and BCG Agentic AI Nov 2025 Report (MIT Sloan and BCG conducted survey of 2,102 respondents representing 21 industries and 116 countries and 11 executives leading AI initiatives in a broad range of companies and industries, including financial services, technology, retail, energy, and health care.)

The report explains, why Agentic AI spreads so fast:

Takeaway is because the infrastructure and user familiarity from GenAI are already in place, agentic features can spread organically without major new deployments. That means competitive edge will come less from “having access first” and more from how well leaders redesign workflows, controls, and incentives around agents.

According to recent McKinsey’s analysis,

Read more

For years, the math was simple: NVIDIA’s CUDA ecosystem was a mandatory cost of entry, securing them estimated 80%+$ gross margins on GPUs and forcing tech giants to pay a massive silicon premium. Today, that economic model is under its most serious challenge yet. The fight has shifted decisively from raw speed to Total Cost of Ownership (TCO), driven by the sobering reality that AI inference costs will soon outweigh training costs by 15x or more over a model’s lifecycle.

To counter this, Google developed the TPU an ASIC built to be ruthlessly efficient. Now, they are weaponizing this externally. The market is currently witnessing Google’s three unprecedented strategic moves that signal a fundamental shift in compute economics:

1. Google’s Balance Sheet Backstop strategy

2. Anthropic Deal

3. Meta Hedge

First, Google is deploying its Balance Sheet Backstop strategy: using its vast financial capacity to guarantee billions of dollars (e.g., a reported $1.3 billion backstop for partners like TeraWulf and Fluidstack) to build new, non-Google data centers.

This move is not charity; it is a calculated financial maneuver designed to solve Google’s two biggest problems in the AI race:

1. Speed-to-Market: Google needs vast amounts of new TPU capacity deployed faster than it can build its own data centers. Partnering with a company like TeraWulf, which already owns power contracts and infrastructure, allows for rapid expansion (e.g., the first phase was expected to come online within months).

2. External Distribution: By placing TPUs outside of GCP through “neoclouds” like Fluidstack, Google lowers the entry barrier for smaller AI companies and developers who might be wary of full GCP vendor lock-in, thus accelerating TPU ecosystem adoption and directly challenging NVIDIA’s installed base.

Google is using its balance sheet to subsidize the CapEx required to stand up an alternative to the NVIDIA/AWS/Azure cloud ecosystem, a move that provides a low-cost, high-efficiency compute option that structurally undercuts the high-margin NVIDIA model.

Second, the foundational model builder Anthropic has made a clear, strategic commitment to buy up to $1 M Google TPUs. This deal, valued at billions, provides critical third-party validation that the TPU offers superior price-performance, delivering up to 4x better cost-performance than NVIDIA GPUs for certain Large Language Model (LLM) tasks.

Third, the threat is magnified by The Meta Hedge: one of NVIDIA’s largest existing customers, Meta, is reportedly in advanced talks to integrate TPUs into its own data centers starting in 2027. This signifies that the cost pressure from TPUs is too significant to ignore, forcing even the most committed GPU users to seek a major hedge and potentially divert billions in future spending away from NVIDIA. The core investment question is no longer about raw power, but about which architecture—the general-purpose flexibility of the GPU, or the specialized, TCO-optimized ASIC backed by Google’s aggressive financial strategy—will secure the lowest perpetual operating cost for the next decade of AI deployment.

We also see venture and investor excitement shifting: Berkshire Hathaway took a stake in Alphabet in 2023/2024, which many interpreted as a bet on Google’s AI strategy Buffett, known for favoring economic “moats,” likely saw Google’s vertical integration as a moat that can challenge Nvidia’s.

We have seen no shortage of blogs and news articles comparing the technical architectures and benchmark performance of TPUs versus GPUs. That is not the purpose of this article. Instead, we are approaching the question from a C-suite perspective, focusing on economic outcomes, capital allocation, and long-term operating leverage—how these platforms change total cost of ownership, risk, and strategic flexibility at scale, rather than which wins a micro-benchmark.

In today’s article, we’ll dig into the questions that have dominated recent client conversations:

1. Simulate a C-Suite Total Cost of Ownership (TCO) discussion between the VPs of Infrastructure/AI at Anthropic (a committed TPU user) and Meta (a large, hedging GPU user). Which accelerator—NVIDIA’s or Google’s —delivers the superior TCO in three distinct scaling scenarios: Base Cluster (1K Accelerators), Medium Cluster (10K Accelerators), and Hyperscale Cluster (100K+ Accelerators)?

2. What is the specific bill of materials (BoM) and supply chain architecture of a modern Google TPU Pod (e.g., Ironwood v7)? How does Google’s vertical integration—from silicon co-development to the interconnect fabric—create an economic moat that provides both significant unit cost advantage and unmatchable supply chain predictability?

3. Software Ecosystem: CUDA vs. XLA – Will Customers Face Lock-In? Can you run Hugging Face (PyTorch) models on TPU? What is the downsides for companies to use TPU vs. Nvidia across the dimensions ?

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

What a “TPU Pod” Actually Is (Rack-Level, Not Chip-Level)

When Google sells TPUs, it is not selling “a chip.”
A TPU Pod is a vertically integrated system in which compute, memory, networking, cooling, and software are designed together and operated as a single economic unit. Thousands of accelerators behave like one machine.

Source: Part of an Ironwood superpod, directly connecting 9,216 Ironwood TPUs in a single domain.

The key to understanding Google’s TPU strategy is that all layers matter, simultaneously. A TPU Pod is not a chip decision; it is a full‑stack system that spans accelerator trays (compute and memory), intra‑rack fabric and host nodes, inter‑rack optical networking, power delivery and direct liquid cooling, and a pod‑level runtime and scheduler that treats thousands of accelerators as one logical machine.

These layers are co‑designed and deployed together, and the economic outcome—$/rack, racks per megawatt, utilization, and speed of deployment—is determined by how well they work as a unified system. Focusing on any single layer in isolation misses the point: Google’s advantage emerges from integrating all layers into a rack‑scale computer with predictable cost and performance at scale.

Let’s look at the layers in detail before we get into TCO scenarios.

Read more

In our previous piece, “The ROI Math Behind Humanoids in Manufacturing,” we unpacked where the economics of humanoid deployment finally start to make sense—how labor substitution, uptime, and CapEx amortization align to create real financial return. That analysis triggered a flood of follow-ups from clients who wanted to go deeper—not into the mechanics or battery chemistry, but into the data economics behind the machines. Investors are now asking sharper questions: What does the robotics data chain actually look like? Who owns it? And what role do players like Scale AI, iMerit, and Encord play in that emerging ecosystem?

Those questions are landing just as production ramps at an unprecedented pace. Tesla expects to produce 500 000 humanoids by 2027, giving it more robots than Japan’s current industrial-robot base.

Figure, Agility Robotics, and BYD are right behind, each targeting between 10 000 and 200 000 units within the next two to three years—effectively creating the world’s first supply chain for human-form automation. At these volumes, even conservative adoption means millions of motion hours per week—a torrent of sensor data no single robotics firm can label or train on alone. That’s why attention—and capital—is shifting toward data-refinery companies like Scale AI, the invisible layer that transforms raw robotic motion into learning fuel.

Zooming out, Morgan Stanley projects the humanoid-robotics market could exceed $5 trillion by 2050, with more than 1 billion humanoids operating across homes, warehouses, and factories.

In the nearer term, Bain & Company’s 2025 Technology Report notes that the sector attracted roughly $2.5 billion in venture capital during 2024. The investment thesis is converging: once robots master reliable movement and dexterity, the limiting factor won’t be metal or cost—it will be data, and who controls the continuous learning loop that turns it into intelligence.

In today’s article, we’ll dig into the questions that have dominated recent client conversations:

1. What is the hierarchy of data in humanoid robotics—embodiment, egocentric, and environmental—and how does each drive model performance?

   1. Which dataset is in higher demand right now—egocentric human-POV video or embodiment tele-operation traces? And is egocentric simply easier to scale?

2. How do robotics companies choose between data infrastructure providers like Scale AI, iMerit, Sama, or Encord?

   1. What a “typical volume contract” looks like for robotics/humanoid-data clients?

3. Who’s actually buying Scale AI’s Physical-AI data services today—and what do they keep in-house vs. outsource?

   1. As humanoid platforms mature, do robotics OEMs keep using Scale AI as an external data partner, or does the role of third-party data providers shrink as companies internalize their own data engines?

4. How analogous are humanoid-robot data pipelines to those of autonomous vehicles?What lessons from AV data collection, labeling, and simulation transfer to Physical AI—and where do the workflows fundamentally diverge?

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Q1. Hierarchy of Data in Humanoid Robotics

Humanoid robotics is moving from polished demos to paid deployments, and the scarce input is no longer cameras or compute—it’s data that reliably turns perception into action on real hardware. For investors and systems integrators, the relevant asset isn’t a single “dataset,” but a hierarchy of training signals that compound: egocentric data to perceive and understand(why behind), environmental (exocentric) data to provide context and safety, and embodiment data to act and adapt on the target robot. This stack determines not only whether a robot can complete a task, but whether it can do so repeatably, safely, and at an economics that scales across sites.

Read more

Power is the rate-limiter. Lawrence Berkeley National Laboratory estimates U.S. data-center electricity rising from 176 TWh in 2023 to ~325–580 TWh by 2028, which is ~6.7–12.0% of total U.S. electricity depending on economy-wide growth.

AI isn’t waiting on the grid. Over the next few years, data-center power demand is set to jump, power utility queues are measured in years, and “optional” levers like liquid cooling, high-efficiency UPS, and on-site generation are becoming the gating factors for shipping compute, not just trimming OPEX. Power has moved from a utility bill to the rate-limiter—and the source of schedule alpha.

In the last few years, Solid-oxide fuel cells programs have scaled from pilots to portfolio strategy. Solid-oxide fuel cells (SOFCs) slot neatly into power plan: they’re modular (1–10 MW blocks), quiet, ultra-low NOx/PM, and can run 24/7 on natural gas, RNG, or hydrogen.

Recently, Equinix disclosed deployments surpassing 100 MW across 19 U.S. facilities; Brookfield announced a $5 B, multi-GW rollout targeting AI data centers; and Oracle expanded collaborations at select U.S. campuses. The pattern is consistent: hyperscalers and colo leaders are using SOFC blocks as schedule insurance—often paired with BESS for ride-through and black-start—so capacity arrives on time while utility work and transmission catch up.

In this piece, we’ll unpack:

1. 10 Key Hyperscaler Power Efficiency Levers

   1. What are the biggest levers to cut power across thermal management, UPS architectures, on-site generation, and control systems—and how quickly can each deploy at scale?

   2. What should a vendor-selection scorecard actually look like?

2. Power Grid Queues and Timelines

   How long does it take, and How do these capacity and bottlenecks differ by region—for example, between fast-moving markets such as ERCOT (Texas) or Southeast non-ISO utilities versus slower, study-heavy regions like CAISO (California), PJM (Mid-Atlantic), and MISO (Midwest)?

3. Trump’s policy changes

   1. What is the recent Oct 2025 policy update? How could Trump’s policy changes regarding data center grid connections? What are the main changes in the stages of the U.S. grid connection process?

   2. Would the policy push hyperscale operators toward building more behind-the-meter generation and storage capacity?

4. Fuel cells in AI campuses

   1. Why on-site power and Where do fuel cells fit for AI data centers?

   2. When do SOFCs beat turbines + BESS on TCO, permitting, and resilience?

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Q1. 10 Key Hyperscaler Power Efficiency Levers

The U.S. Energy Department’s 2024 data-center energy report says the sector’s electricity use is accelerating: about 7% CAGR from 2014–2018, jumping to 18% in 2018–2023, and projected at ~13–27% annually through 2023–2028.

Source: 2024 United States Data Center Energy Usage Report

Let’s quickly look at data center electricity projections by equipment types:

Read more

Lately, we’ve been getting a surge of questions about Serverless GPUs — what they are, how they differ from traditional GPU-as-a-Service, and most importantly, what the ROI looks like. Across our client base, there’s growing demand for quantitative frameworks that compare serverless GPU deployments against always-on clusters, measuring utilization-driven savings and time-to-scale advantages.

When we published our CoreWeave analysis back in May, the stock was trading around $45. Since then, it has soared by 168%. In that same piece, we also examined the broader GPU-as-a-Service (GaaS) market and CoreWeave’s strategic deal with OpenAI — one of the first signals that GPU infrastructure itself was becoming a profit center, not a commodity.

That thesis is now being validated across the ecosystem. In September, Microsoft signed a $17.4 billion agreement with Nebius to secure long-term GPU capacity for its AI operations — with upside provisions taking the total closer to $19.4 billion. Markets reacted instantly: Nebius stock jumped nearly 47% pre-market on the announcement.

Meanwhile, on the inference side, Together AI has been on a rapid climb. The company now carries a $3.3 billion valuation, reports $50 million in 2024 revenue, and is guiding toward $120 million for 2025 — proof that serverless inference APIs can scale into real enterprise revenue.

All of this is reinforcing the CAPEX forecasts we’ve seen from Citi and others, who continue to project a multi-trillion-dollar AI infrastructure super-cycle. Citi recently raised its estimate for cumulative AI-related infrastructure spending by major tech companies to $2.8 trillion through 2029, up from $2.3 trillion earlier this year. The same report forecasts AI revenues hitting $780 billion by 2030, up from $43 billion today — underscoring how under-modeled this sector still is.

As Citi’s analysts put it:

“Our conversations with CIOs and CTOs reflect a similar increase in urgency around adoption at the enterprise level. Based on this, we are raising our 2026E forecast for AI CAPEX across hyperscalers from $420 billion to $490 billion, and through 2029 from $2.3 trillion to $2.8 trillion.”

Those forecasts align closely with new data from Dell’Oro Group, which shows CAPEX growth across AI data centers accelerating at a double-digit CAGR well into the next decade.

So while “Serverless GPUs” might sound like a close cousin of “GPU-as-a-Service,” it’s actually a different operating model — with its own economics, scaling behavior, and cost levers. That’s where we’ll focus today’s piece:

1. Operating Model Deep Dive: Serverless GPU vs. GPU-as-a-Service

2. Benchmarking the Serverless GPU Landscape

   1. Economic KPIs – pricing, utilization efficiency, and unit cost

   2. Technical / Operational KPIs – latency, cold-start performance and scaling

3. Use Case Spotlight: How Substack Leverages Modal AI

4. ROI Calculator for Serverless GPU Migrations: A structured framework to estimate cost savings from migrating to serverless GPU environments, using two modeled workloads:
   a. Case A — Small / Spiky Load (Modal Serverless)
   b. Case B — Large / Steady Load (Together AI Dedicated / Hybrid Serverless)

We’ve sketched the economics of serverless GPU adoption into a hands-on Excel model that lets you simulate your own workloads. The tool compares traditional provisioned GPU costs with serverless pricing, accounting for utilization, concurrency, and idle overhead.

You can tweak sliders for GPU hours, utilization %, and burst frequency to see how the break-even point shifts between Case A (small, spiky workloads) and Case B (large, steady inference pipelines). The output updates instantly to show total monthly cost and ROI making it easy to visualize the financial impact of going serverless before running a single experiment.

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Read more

After Meta’s 49% stake in Scale AI, investors are increasingly asking what other adjacent AI data service sub-segments could drive growth or margin expansion. Meta stake has sharpened focus not only on neutrality and concentration risk but also on the pathways for Scale AI and other companies to extend beyond its current core and build a broader moat in the evolving data-for-AI ecosystem.

The last 12–18 months have delivered some signals. The open-web free-for-all is over. As we noted in our previous article, OpenAI’s landmark deals with News Corp are now a strategic moat. Budgets are moving from ad‑hoc red teaming to programmatic evaluation‑as‑a‑service with regulator‑ready audit trails. In a recent development, Anthropic and the U.S. government's nuclear experts have developed a new tool that can spot the difference between a scientist asking Claude about nuclear reactors and a spy probing it for secrets about weapons development.

OpenAI and the UK government have signed strategic partnerships.

The rise of agentic AI—autonomous, tool-using systems like Microsoft Copilot agents and OpenAI Realtime agents—is changing what counts as “gold” in this industry: traceable interaction data and multi-turn evaluations.

We’ll also take a closer look at near-term plays from Scale AI and Innodata.

With all of these developments shaping the market, today’s article addresses the questions we are hearing most frequently from clients and investors:

1. Who’s really buying AI data today?
   Review 2024 State of AI Reports findings and understand market dominance for data

2. Which services matter most?
   RAG surge data & within the AI data services stack, what sub-segments are proving most interesting to clients and investors?

3. Who’s raising capital?
   Several AI data companies have raised significant rounds recently. Which ones stand out, and what’s the equity story behind them?

4. What’s the short and long-term trajectory?
   What does “forward-deployed capability” actually look like in practice?

   What long-term shifts should we expect beyond forward-deploy?

5. What is the growth outlook for the mid-market segment spending $50K–$500K annually?

   Review DataBricks AI Report and understand Scale AI’s opportunities

6. What are platform-control, on-premise trends?

   1. Are enterprises pushing to use their own annotation platforms, or are they adopting proprietary data platform?

   2. Are clients asking for on-premise annotation teams, embedded experts, or dedicated secure centers instead of purely cloud-based delivery?

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Q1. Who’s really buying AI data today?

According to survey data from the CIO vision 2025 report by MIT Technology Review Insights, 72% of leaders say that problems with data are significant.

Read more

Beyond headline names like Nvidia and Tesla, investors are beginning to map out the broader humanoid robotics landscape. In recent months, we’ve spoken with LPs and public-market funds who acknowledge it may feel early as a 2025 investment theme—the common view being that true scale manufacturing is still a couple of years out. Yet a growing group is al…

Read more

In our recent deep-dive, we highlighted that many investors are eager to understand Elastic’s offering in greater detail. Elastic is back in the spotlight because its search-first architecture now underpins high-growth arenas like GenAI retrieval, vector search, and hybrid observability-security stacks

In a recent Elastic’s earning call, CEO described an eight-figure expansion deal with a global financial institution and a multi-year seven-figure expansion with an international banking group.

"We expect total revenue in the range of $1.655 billion to $1.67 billion. Representing 12% year-over-year growth at the midpoint or 11% year-over-year constant currency growth at the midpoint."

"We announced a new five-year strategic collaboration agreement with AWS, which deepens our partnership through solution integrations, joint go-to-market, and marketing initiatives."

Given Elastic's expanding role across full-text, semantic, and operational search layers, investors are conducting a targeted set of diligence conversations to better understand how the platform is being adopted in practice — including areas where it displaces incumbents or complements cloud-native data services.

Elastic operates at the intersection of several high-growth segments — including enterprise search, log analytics, observability, and increasingly, vector search and RAG architectures.

The global enterprise search market is expected to grow at a CAGR of ~10–13% through 2030, with adjacent markets like observability and vector search expanding even faster. As spending shifts from batch-oriented data lakes to real-time and AI-native systems, Elastic’s open-source foundation, scalable architecture, and recent investments in semantic search make it a strategic asset under evaluation by both technical teams and procurement stakeholders.

Many of our clients have expressed heightened interest in Elastic’s Application Search business, so this article concentrates primarily on that segment. We’ll also provide a brief overview of Elastic’s competitors in security and observability.

Let’s start with the questions from clients:

1. Why is Elastic popular and which market segments does it most effectively address?

2. Elastic’s competitive positioning in Application Search, Observability and Security

   1. How does Elastic’s App Search Business compare against competitors such as OpenSearch, Glean, and Coveo?

      1. TCO high level cost model with Baseline and Active Analysis scenario between Elastic and OpenSearch

3. When choosing a storage and search solution for Generative AI and “Agentic AI” workflows, what several factors come into play?

4. With a shift underway from GenAI interfaces to more autonomous, agent-As enterprise data stacks adapt for generative-AI and agent-based workloads, how well is Elastic positioned to capture this opportunity, and what specific factors support that assessment?

5. Did Elastic “reverse” its 2021 license switch? What are the impacts? Does this spill into Observability and Security?

6. Self-hosting and open-source drift ― why it’s a real risk for Elastic?

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Read more

We know we promised our next piece would cover Elasticsearch—but after a strong wave of inbound questions on Celestica and its role in GenAI infrastructure, we’ve decided to bump it up in the queue. Don’t worry—our deep dive on Elasticsearch is still coming next.

Lately, a lot of our clients have been asking about Celestica—how a company known for low-margin hardware is suddenly showing up in hyperscale AI supply chains. A 99% jump in networking segment revenue reflects more than just cyclical tailwinds—it’s a sign that Celestica is increasingly being pulled into core AI infrastructure design wins.

While historically known as a low-margin ODM, Celestica is now emerging as a critical enabler of hyperscaler AI buildouts, particularly in white-box switching and rack-scale delivery. Clients are asking sharp questions: What’s their moat? Why do Meta and Microsoft buy from them? Can they scale beyond just ToR (Top Of Rack) switching into core spine/leaf switching AI fabrics where traditionally Arista is the leader?

Meta, one of Celestica’s largest AI infrastructure customers, has publicly stated that the future of AI networking at hyperscale data centers will be built on disaggregated, white-box platforms.

This isn’t just a philosophical endorsement of open architectures—it’s a procurement directive. Meta’s own Disaggregated Scheduled Fabric (DSF), which underpins its GenAI training clusters, is built from switches that are based on Broadcom silicon, supplied by partners like Celestica. This validates Celestica not just as a vendor, but as a core execution partner in Meta’s long-term AI fabric roadmap.

In today’s article, we’ll break down how the GenAI networking stack is evolving—and why Celestica is increasingly viewed as a core beneficiary of this shift. We’ll wrap up this blog with a directional revenue model that quantifies what Meta’s AI cluster buildout could mean for Celestica’s switch business—based on public topology data, conservative ASPs, and realistic share assumptions.

Topics include:

1. GenAI Network Topology & Meta’s Reference Architecture

   1. Comparing Three Network Models: We contrast Chassis-Based Routers, traditional Spine–Leaf, and Distributed Disaggregated Chassis (DDC) to highlight the architectural evolution.

2. Inside Meta’s DSF: The 4K and 16K GPU Pod Topologies: A dive into Meta’s Disaggregated Scheduled Fabric, including VOQ scheduling, rack-scale design, and vendor implications.

3. AI Training vs. Inference Data Centers: We explain why training and inference have fundamentally different networking requirements—and what that means for infrastructure buyers.

4. The Outlook for White-Box, Disaggregated Switching: CAGR, Market projections and how open platforms are gaining traction

5. Celestica’s Four Strategic Moats (and Who They Compete With)

6. Directional Revenue Model: What Meta’s AI Cluster Build Means for Celestica: Using Meta’s own topology disclosures, we estimate how many switches are involved and how much revenue could flow to Celestica per generation.

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Read more

In our recent client conversations—we're seeing a surge of interest in two platforms in particular: Snowflake and Elastic. Once seen as complementary or niche, both are now being evaluated as foundational layers for AI readiness, cloud spend visibility, and workload consolidation. The questions enterprises are asking today signal a shift: from “How do we store data?” to “How do we operationalize it—and at what cost?”

Once viewed as back-office plumbing, the cloud data warehouse has evolved into the strategic control plane of the modern enterprise—underpinning everything from AI model pipelines and real-time analytics to governance, compliance, and workflow automation. But few companies embody this transformation more fully than Snowflake.

Snowflake is no longer just a high-performance SQL engine. With the rollout of its Generation 2 (Gen2) Warehouses, the launch of Openflow (a fully managed data integration engine), and the maturation of its Cortex AI platform, Snowflake is positioning itself as the core runtime for enterprise-grade GenAI.

This article dives into Snowflake, offering a glimpse into the real questions clients are asking—exploring the infrastructure moves, Cost, AI-native design decisions, and monetization layers that investors and operators alike are closely tracking.

In Part 2, we’ll turn to Elastic, a platform long associated with observability and enterprise search, but now expanding its role in AI-powered logging, vector search, and real-time semantic analytics.

Let’s start with the sharpest questions from clients shaping enterprise Snowflake adoption today:

1. Cloud Data Warehousing Trends

   1. What is the current size and projected CAGR of the cloud data warehousing market?

   2. How do Snowflake, Databricks, and BigQuery compare across the core layers of GenAI infrastructure?

2. Snowflake – Cost models, Workloads, Expansion, and AI Relevance

   1. Why are only around 40% of Global 2000 companies using Snowflake, despite its leadership in cloud data platforms — and does the continued presence of Redshift, Fabric, BigQuery, Databricks, and legacy players like Teradata/Dremio/Oracle suggest that the market is already mature, or is broader enterprise adoption still ahead?

   2. What is Snowflake's new feature Generation 2 Engine? & What is Snowflake's Openflow?

   3. How does Reverse ETL adoption trends benefit stickiness on Snowflake?

   4. How does Snowflake’s consumption cost model scale across small AI teams, GenAI Ops units, and enterprise-grade AI infrastructure?

   5. How are organizations splitting Snowflake usage between mission-critical workloads and exploratory analytics or ML?

   6. In a tighter budget climate, how are enterprises optimizing Snowflake usage without sacrificing agility or performance?

   7. Is Snowflake growth being driven by more users—or by heavier usage per user across AI and analytics teams?

   8. Is Snowflake still “just a business intelligence data warehouse”—or is it becoming a full-stack transactions(OLTP), analytics(OLAP) and vector database platform to support the emerging gen AI applications ?

   9. Why growing number of enterprises have both Snowflake and Databricks deployments?

   10. Does zero-copy ETL with Salesforce Data Cloud represent a strategic synergy—or a potential headwind for Snowflake compute revenue?

3. ElasticSearch (Part 2)

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Cloud Data Warehousing Trends

We’ve seen a noticeable shift in how enterprise data teams think about their warehouses—not just as reporting layers, but as engines for AI enablement. The demand curve is no longer tied just to BI headcount—it’s expanding across ML, marketing ops, and even product teams

Enterprises continue to migrate data warehousing to the cloud for flexibility and scalability. In recent surveys, only about 18% of IT managers still keep all data warehouses on-premises, while roughly 47% have moved entirely to public cloud and the rest use a hybrid mix.

Read more

Between 2022 and 2025, licensing data for AI model training evolved from a niche practice into a standard industry paradigm. In 2024, Reddit’s COO noted that Reddit’s agreements with Google and OpenAI already comprised ~10% of the company’s revenue, underscoring how significant this new revenue stream became for data-rich platforms.

At the high end, the News Corp–OpenAI deal (announced in 2024) is reportedly worth more than $250 million over five years. OpenAI didn’t just buy headlines—they licensed revenue retention. The News Corp deal marks a new era where quality content becomes recurring margin in the AI stack

Another one,

That $16M is only the fixed minimum – the deal includes a variable component that could boost payments further based on how the content is used.

As of now, most leading AI firms had inked multi-year data agreements with content providers, and many publishers saw licensing as a material revenue line. In short, a “data sharing economy” has emerged, in which content owners monetize their archives and real-time data, while AI companies secure the reliable, proprietary datasets needed to train and enhance AI models.

Later in this post, we will break down a directional revenue model for TIME’s partnership with Perplexity.ai—one of the first AI-native search engines offering ad revenue sharing to publishers. While the numbers are built on public traffic estimates and industry benchmarks, they offer a pragmatic lens into how premium publishers could begin to monetize their content in AI-driven environments.

It’s not perfect math, but it’s a solid directional framework for investors looking to understand where the upside lives in this evolving content stack.

This article will go through the AI Companies and Media Partnership landscape, breaking down:

1. AI Data & Content Ecosystem & Players

   1. Most Valuable Types of Data for GenAI Training

   2. Breakingdown players in three categories: 1 . Scarce upstream IP rights players 2 . Players benefit margins from expert labor & labeling tech & 3 . Toll booth and audit-as-a-service players

2. Why Content Originators Are Becoming Critical Infrastructure

   1. Rise of licensing deals (News Corp, Reddit, etc.)

   2. Breakdown of which AI companies are aligned with which publisher

3. Buying the Moat: AI–Media Licensing Model is Maturing in 2025

   1. Various Contract structures & pricing

   2. OpenAI & NewsCorp contract structure

4. Who Gets Paid When AI Scales?

   1. Attribution-driven traffic + rev share (e.g., Perplexity.ai’s model) is getting traction

   2. Perplexity–TIME revenue uplift analysis

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

AI Data & Content Ecosystem & Players

While the generative AI boom has triggered massive investment into model development, it's important to distinguish between two fast-growing but fundamentally different markets: AI Data and Content licensing and AI training/fine-tuning dataset spend. The data licensing market refers specifically to payments made by AI companies to content owners—such as publishers, platforms, and forums—for legal access to their proprietary text, images, or metadata.

In contrast, training and fine-tuning dataset spend captures the entire end-to-end cost of making data usable for models. This Market includes companies providing data annotation, curation, synthetic data, AI model training, and evaluation. We expect this broader dataset operations market could surpass $30 billion annually by 2030, particularly as enterprises fine-tune foundation models for domain-specific use.

We will look at AI Data and Content Market for this article.

To understand how the generative AI economy is forming around data, we need to first map the emerging ecosystem of content and infrastructure players shaping the flow of information into foundation models.

The AI content ecosystem is evolving into a structured value chain, with distinct upstream and downstream roles. At the foundation are 1. Content & Signal Originators—publishers, platforms, and API providers whose data powers generative models through training or retrieval. Sitting above them are 2. Data Enrichers like Scale AI and Surge AI, which convert raw content into high-quality training assets via labeling, alignment, and filtering.

Finally, 3. Licensing Infrastructure players—like Prorata or API brokers—are emerging as gatekeepers, enabling monetization, compliance, and rights orchestration. For investors, this stack highlights where durable moats, pricing power, and second-order data royalties may emerge in the GenAI economy.

For hedge funds and private equity firms, the maturing AI content ecosystem creates investable opportunities across all three layers of the stack. At the top, scarce content rights holders—like publishers, Q&A platforms, and financial APIs—are turning into recurring revenue engines through licensing, making them prime for cash flow extraction or roll-up strategies.

In the second layer, data enrichers and RLHF platforms like Scale AI, Appen, and emerging synthetic data startups offer high-margin, defensible infrastructure plays—essential for fine-tuning and safety alignment in models. Finally, the third layer—licensing and monetization infrastructure—represents a greenfield opportunity. Firms like Prorata.ai are creating the plumbing for data revenue flow, much like ASCAP or BMI did for music, and may evolve into tollbooth-like assets in the GenAI value chain. For investors, the play isn't just the model builders—it's the upstream and infrastructure providers who will extract value every time a model trains, responds, or monetizes.

When an AI model company is looking to license or buy data, they don’t just ask, “How much data is available?”—they ask, “What kind of data will make our model meaningfully better for specific use cases?” The goal isn’t raw scale anymore—it’s strategic data that improves output quality, safety, factual grounding, and domain-specific reasoning.

For general-purpose LLMs, companies seek high-quality language data that helps the model write more fluently, reason better, and minimize hallucinations. This often includes well-edited, factual, and stylistically diverse text: think journalistic content, encyclopedic entries, academic papers, and books with narrative structure. But as the market matures, the focus is increasingly shifting to specialized and proprietary datasets—especially enterprise content.

One of our clients asked: What types of data are considered most valuable for AI companies to acquire or license—and how do those companies evaluate what’s worth paying for?

Read more

2024–2025 saw major deals in GPU-as-a-Service space. OpenAI’s deal with CoreWeave was one; another was Amazon’s $4 B investment in Anthropic (OpenAI rival) which bound Anthropic to primarily use AWS for training and even integrate Amazon’s own AI chips​. Some recent moves:

Over the past several weeks, we’ve seen a surge in inbound interest from clients—ranging from hedge funds to AI infrastructure teams—wanting to understand CoreWeave’s position in the GPU-as-a-Service landscape.

As CoreWeave prepares to release its Q1 2025 earnings on May 14, market attention is shifting from curiosity to conviction. The company went public on March 28, 2025, listing on the Nasdaq under the ticker symbol CRWV. The IPO raised $1.5 billion by offering 37.5 million shares at $40 each, resulting in an implied valuation of approximately $23 billion and a multi-billion-dollar compute deal with OpenAI already inked, CoreWeave is emerging as one of the most closely watched companies in the generative AI infrastructure ecosystem.

The question is no longer “What is CoreWeave?” but “Where do hyperscalers fall short—and CoreWeave wins?” Clients evaluating high-throughput workloads are starting to recognize that general-purpose clouds aren’t always the best fit for specialized AI needs. For many investors and operators alike, CoreWeave’s combination of NVIDIA-first supply access, AI-native infrastructure, and pricing that undercuts hyperscalers is raising a bigger question: Is GPU cloud the next great infrastructure opportunity—and is CoreWeave the purest way to play it?

According to recent KPMG report on GPU investment, the result consists of a survey of 120 highnet-worth individuals (HNWIs) and investment professionals – indicated that a significant cohort of market participants have already gained exposure to the GPU market (78% of those surveyed).

It appears the KPMG survey didn’t include a distinct “GPU-as-a-Service” category, but the rising interest in GPU infrastructure funds, direct GPU manufacturers, and AI companies reliant on GPU tech suggests that investors are actively circling this space. In many ways, GPUaaS is still flying under the radar as a formal asset class—but for early movers, that’s exactly where the opportunity lies.

Per KPMG, GPU focused ETFs are the preferred vehicle of choice, with 71% of advisers and 62% of high-net-worth individuals (HNWIs) investing in them in recent years.

In this article, we will review:

1. Understanding the GPU Market Landscape

   1. Defining GPU-as-a-Service (GPUaaS)

   2. Current GPUaaS Market % split by Use Caes: AI training, inference, cloud gaming, media/visual effects

2. Market Players: Who’s Powering the GPU Cloud?

   1. CoreWeave vs Hyperscalers: Who Wins When?

3. Inside CoreWeave’s Moat: Why it Isn’t Just Another GPU Provider

   1. Breakdown of how CoreWeave stacks up against Lambda Labs and major Hyperscalers

   2. What is CoreWeave’s LOTA? Does it really matter in Training vs. Inference?

   3. Pricing: Simple, Transparent, and Built for AI

4. Case Studies: Strategic Choices in the Real World

   1. OpenAI’s multi-billion dollar deal with CoreWeave

   2. Cohere’s selection of CoreWeave for next-gen LLM infrastructure

   3. xAI’s transition from cloud to custom GPU data center buildout

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Understanding the GPU Market Landscape

The Global GPU market is estimated at around $87B in 2025, and expected to hit $365B by 2030, at a CAGR of over 33% . Moreover the report mentions a substantial shift toward GPU as a service model.

Read more

Meta’s blog on “A case for QLC SSDs in the data center” is getting a lot of traction, and our clients are asking pointed questions: Is QLC really ready? Will Pure Storage win big? What happens to HDDs? Our clients are interested in learning more about Data Storage for Cloud and AI and companies like Vast, Weka, FlashBlade, and others similar.

Now, if you…

Read more

At GTC 2025, Nvidia emphasized that we are at an inflection point in computing, where AI data centers are becoming the new factories of the digital age. CEO Jensen Huang highlighted that Data Center CapEx is projected to exceed $1 trillion by 2028, reflecting the massive investment into AI infrastructure by hyperscalers.

Source: GTC NVDA 2025

While DeepSeek has shown that AI models can be trained with fewer GPUs and at lower cost, one might expect this to slow the overall growth in AI compute demand. However, the data tells a very different story.

According to another 650 Group report, companies spent over $200 billion on AI hardware in 2024—an all-time high—driven by the rapid adoption of generative AI, hyperscaler expansions, and enterprise AI deployments. This raises an important question: If DeepSeek's advancements in model efficiency were meant to reduce GPU dependency, why is AI hardware spending still skyrocketing?

In a world where AI models are becoming more sophisticated, multimodal, and reasoning-driven, clients are asking critical questions: How much compute is enough? More specifically,

• How many GPUs are being used for inference? Are custom AI chips (TPUs, Trainium, Maia AI) reducing dependence on Nvidia hardware?

• What does the 2025–2027 roadmap for AI compute look like across hyperscalers?

This blog explores the post-DeepSeek demand landscape, breaking down:

1. Contrasting the Pre-DeepSeek vs. Post-DeepSeek

   1. Why Efficiency Gains From DeepSeek haven’t Slowed the AI Compute Boom?

2. AI Compute Deployment Forecast by players: GPUs & ASICs (2024–2026E)

   1. Players: Microsoft,Google, Meta, Amazon, xAI (Elon Musk startup)

3. AI Inference Demand Forecast By Use Cases (2024–2027E)

   1. Use Cases: Chatbots & Assistants, AI Agents, Video & Image AI, Enterprise AI and Autonomous Systems

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Read more

The data center industry has been sent into a frenzy with recent reports that Microsoft has canceled select leases with at least two private data center operators in the U.S., totaling "a couple of hundred MWs." According to TD Cowen, Microsoft’s cancellations involve leases with at least two private data center operators, and Microsoft has also pulled back on converting Statements of Qualifications into formal leases.

Our clients have been asking the question: Who was the lessor left holding the bag after Microsoft abruptly canceled hundreds of megawatts in data center leases? Industry whispers point to a handful of key players that we will review.

Let’s go back a month before this lease cancellation.

In January 2025, Microsoft said it plans to spend $80B on building AI data centers in the 2025 financial year, The $80B would reflect a significant increase on the $53B CAPEX spend Microsoft made in 2023. Below comes officially from Microsoft:

Two weeks ago, Microsoft CEO Satya Nadella sat at an interview in an hour-long session.

When asked for data center investments, Nadella acknowledges that some people compare the current data center expansion to a "bubble," similar to the 1990s dot-com era. However, he argues that the infrastructure built during that time laid the foundation for the modern internet. He implies that the current AI-driven data center expansion is likely a long-term, foundational investment rather than a fleeting trend.

He emphasized that Microsoft is not just blindly expanding compute capacity but is carefully balancing infrastructure investments with actual AI workloads.

After this interview comes the news of lease cancellation.

Given all this, our clients are wondering what’s really happening behind the scenes. Is this merely a product of supply chain constraints, or is it evidence of a deeper strategic shift—perhaps involving DeepSeek or a more extensive partnership with Oracle?

In this article, we will discuss:

1. The Race for AI Compute: Global Data Center Demand Through 2030

   1. Goldman Sachs, Gartner and McKinsey Report Review

2. Microsoft’s Surprise Pullback: Why the Lease Cancellations Sent Shockwaves

   1. Understand Data Center Lease Staging: SOQ, 500 docs and Final Sign

3. The DeepSeek Theory: A Game-Changing AI Shift or Mere Speculation?

   1. The Oracle Hypothesis: A Shift in AI Cloud Alliances?

   2. Who Was Left Holding the Bag? : Likely Lessors

4. The Supply Chain and Power Bottleneck Narrative

5. A Domino Effect? What This Means for Other Hyperscalers

Read more

As we progress toward AGI, it is becoming increasingly evident that data quality will have a greater impact on model performance than algorithmic advancements alone. However, this fundamental shift is not yet fully understood in the AI research community.

This article explores how data quality impacts model performance. We examine evidence supporting data’s rising importance over model architecture, GPU compute capacity and algorithms ; the expansion of the AI data services market, emerging benchmarks such as Humanity’s Last Exam (HLE) and the transition toward reinforcement learning (RL) and expert feedback in model training.

This growth is reminiscent of the cloud computing boom a decade ago or the big data analytics surge – as AI adoption grows, ancillary services supporting AI (like data labeling, MLOps, etc.) explode. Just as SaaS/cloud became essential for businesses undergoing digital transformation, data services are becoming core to any AI-driven transformation. Companies that once managed data in-house now increasingly outsource to specialists or platforms to keep up with volume and quality needs.

Furthermore, Our clients are probing on Deepseek’s pure RL data approach and its implications on AI data service market. In Question 4 of our last DeepSeek Post, we looked at what DeepSeek’s papers are saying about reasoning samples and SFT samples. On Page 3 of DeepSeek paper, it talks about taking pure RL approach without supervised data. Taking a purely Reinforcement Learning (RL) approach is genuinely revolutionary.

In this article, we will discuss:

1. Analysis of Two McKinsey Gen AI Reports and AI Data Services TAM

2. Model Benchmarking & Eval

   1. Scale AI’s Humanity’s Last Exam (HLE) And Performance Of The Latest AI Models

3. Emerging Gen AI Data Ecosystems And Their Impact On AGI

   1. Why Is Supervised Fine-Tuning (SFT) No Longer Enough?

   2. Rise of RL and RLHF for AI Training Are Shaping More Intelligent AI

   3. What is Agent Data and How Does It Differ from RLHF?

   4. Synthetic Data: NVIDIA’s Nemotron-4 340B

4. Top Companies Market Leaders

Read more

We Have Been Inundated

We've been handling a continuous stream of expert network calls since Jan 16th 2025, as everyone is scrambling to analyze DeepSeek’s developments and their far-reaching implications. There have been also deep internal discussions at the companies we work for as well and DeepSeek is being deployed internally and evaluated already. After deeply reading the DeepSeek-V3 and DeepSeek-R1 papers, and discussing this in-depth in our calls, we will give our views on the common questions asked.

Key Questions from Our Clients That We Will Address Here

1. GPU vs. ASIC Distribution: How will the balance between GPUs and ASICs evolve in the coming years for both AI training and inference? What are the expected shifts in hardware allocation across different AI workloads?

2. AI Training vs. AI Inference Data Centers: How will the industry differentiate between training-focused vs. inference-optimized data centers? What will be the impact on capital expenditures (CapEx), and which hardware and infrastructure components (e.g., networking, cooling, power) will be most affected?

3. Industry Impact: Which companies and sectors will feel the greatest impact from DeepSeek’s advancements? This includes key players such as:

   • NVIDIA

   • Vertiv (data center infrastructure and power solutions)

   • Hyperscalers (AWS, Google, Azure, Meta, Tencent, etc.)

   • MongoDB/Elasticsearch

   • Training Data providers (Scale AI, Innodata, etc)

   • Synopsys/Cadence (impact on AI-driven chip design and verification)

4. Synthetic Data vs. Human Data: As AI shifts toward self-learning models and reinforcement learning, what role will synthetic data play in future training processes? How will the data need evolve: Supervised Fine Tuning (SFT) data vs. Reinforcement Learning (RL) data?

5. Edge AI vs. Cloud AI: Will AI inference remain centralized in the cloud, or will advancements in hardware efficiency and model distillation push AI inference closer to edge devices? How will the economics of cloud-based vs. on-device AI evolve in the next 3–5 years?

6. How Did DeepSeek Outperform o1 at a Fraction of the Cost?

   • How was DeepSeek able to scale so efficiently while Western teams struggled with higher costs and longer development cycles?

   • What architectural and training innovations allowed them to leapfrog existing approaches?

7. MoE vs. Dense Architectures: What are the trade-offs between Mixture-of-Experts (MoE) and Dense models in terms of efficiency, scalability, training cost, and inference performance? How will this influence the future direction of AI model architectures?

8. Rights of Use: What are the legal, ethical, and competitive implications of AI companies using outputs from closed-source models to improve open-source alternatives?

Please note: The insights presented in this article are derived from confidential consultations our team has conducted with clients across private equity, hedge funds, startups, and investment banks, facilitated through specialized expert networks. Due to our agreements with these networks, we cannot reveal specific names from these discussions. Therefore, we offer a summarized version of these insights, ensuring valuable content while upholding our confidentiality commitments.

Lets go through these questions one by one:

1. GPU vs. ASIC Distribution

Given, DeepSeek’s success, a central question is how the mix of GPUs vs. ASICs will evolve for both AI training and inference.

Read more