AI and Investing Is the Future!
A quick note: because of the payment processor issues some of you hit when trying to subscribe (especially Apple Card/Apple Pay), I wanted to put out one last free piece and give everyone more time to sign up with the early-bird discount before GTC kicks off Monday. A few subscribers asked me to go deeper on the NeoCloud hypothesis. So here it is. Enjoy.
The discount is extended through end of day Friday, March 13th: bepresearch.substack.com/vip
When NVIDIA ships Rubin later this year, who installs it first?
Not Amazon. They’re juggling Trainium3, Trainium4, and a $10 billion custom chip business growing triple digits. Not Google. They just shipped TPU v7 Ironwood and signed Anthropic to a million-chip deal worth tens of billions. Not Microsoft. Azure is deploying Maia 200 for inference while managing the OpenAI relationship and a million enterprise customers.
The answer is CoreWeave and Oracle. They deploy first because they have no competing silicon programs, no legacy workloads to protect, and no reason to slow-walk NVIDIA’s latest hardware. They are NVIDIA’s distribution layer, the companies that turn Jensen’s silicon roadmap into available compute faster than anyone else.
This is the NeoCloud hypothesis. And understanding it, including its risks, may be the most important infrastructure investment framework for 2026.
The term gets thrown around loosely, but I define it precisely: a NeoCloud is a cloud infrastructure provider whose primary business model is deploying NVIDIA GPUs at scale, without the distraction of competing silicon programs, legacy enterprise migrations, or adjacent business lines that consume management bandwidth and capital.
By this definition, the field narrows quickly. CoreWeave is the purest expression: a company that went from crypto mining to the largest independent GPU cloud in under five years, with $66.8 billion in revenue backlog and NVIDIA as both its primary supplier and strategic investor. Oracle is the more complex case: a 49-year-old enterprise software company that Larry Ellison has aggressively repositioned as an AI infrastructure powerhouse, with $553 billion in remaining performance obligations and FY27 revenue guidance just raised to $90 billion.
Everyone else is something different. AWS, Azure, and GCP are hyperscalers. They build custom silicon alongside NVIDIA deployments, serve millions of enterprise workloads beyond AI, and make capital allocation decisions across dozens of competing priorities. They matter enormously for the AI buildout, but they’re not NeoClouds. They’re conglomerates with AI divisions.
Oracle made this distinction explicit. In Q2, Ellison announced the sale of Oracle’s stake in Ampere, the ARM-based chip company: “We no longer think it is strategic for us to continue designing, manufacturing and using our own chips in our cloud datacenters. We are now committed to a policy of chip neutrality where we work closely with all our CPU and GPU suppliers.” That’s the NeoCloud hypothesis in one sentence. While hyperscalers split capital and engineering bandwidth across competing silicon programs, Oracle chose to be the fastest deployer of whatever chips its customers want to buy. The contrast with Amazon’s Trainium, Google’s TPU, and Microsoft’s Maia couldn’t be sharper.
Here’s the dynamic most investors miss: in a supply-constrained environment, speed of deployment is the competitive moat. But it’s specifically speed of NVIDIA deployment that matters for the NeoCloud hypothesis, and the distinction is becoming more important as hyperscalers pour billions into competing silicon.
When NVIDIA announces a new architecture (Blackwell last year, Rubin this year), there’s a finite number of systems available in the first 6-12 months. The cloud providers that deploy fastest get the first customers, lock in the highest-margin contracts, and build operational expertise that compounds with each generation. Google Cloud will be among the first to offer NVIDIA’s Vera Rubin platform, but Google’s primary capital allocation is flowing to its own Ironwood TPUs. Amazon just announced Trainium4 with NVLink Fusion support. They’re actively building hybrid NVIDIA/Trainium clusters rather than going all-in on NVIDIA racks. Microsoft’s Maia 200 is already deployed and running GPT-5.2.
This is the structural opening for NeoClouds: the hyperscalers are increasingly hedging their NVIDIA exposure with custom silicon, which means their pure NVIDIA deployment speed is diluted by competing priorities. CoreWeave and Oracle have no such dilution.
CoreWeave has demonstrated this repeatedly. They achieved industry-leading MLPerf benchmark results, meaning they don’t just install GPUs; they optimize the entire software stack around them. Their bare-metal approach strips away the virtualization overhead that slows traditional clouds, giving AI labs direct hardware access months before hyperscalers complete their broader integrations.
Oracle has shown similar agility. They delivered close to 400 megawatts of data center capacity in a single quarter and increased GPU capacity by 50% quarter-over-quarter. OCI revenue just hit $4.9 billion in F3Q, up 84% YoY, with a run rate approaching $20 billion annually. FY27 revenue guidance was raised to $90 billion. Jefferies noted that Oracle has reduced rack-to-revenue time by 60% over the past few months, meaning they’re converting installed capacity to paying customers faster than ever. For a company that five years ago was written off as a legacy database vendor, that trajectory is remarkable.
Contrast this with the hyperscalers. Amazon is simultaneously ramping Trainium3 to full commitment by mid-2026, managing the world’s largest logistics network, and navigating a retail business with razor-thin margins. Every dollar of AWS capex competes with fulfillment center investments. Google is pouring capital into TPU v7 Ironwood, a chip that SemiAnalysis says delivers 44% lower TCO than GB200, consuming engineering bandwidth and data center floor space that could go to NVIDIA GPUs. Microsoft’s Azure must balance capacity across enterprise customers, Maia 200 deployment, the OpenAI partnership (which has its own contractual compute commitments), and an emerging Copilot ecosystem that demands inference capacity internally.
These are all great businesses. But they have structural friction that prevents them from being NVIDIA’s fastest distribution partners. It’s not a question of competence. It’s a question of competing priorities.
At Deco Lighting, I saw this pattern play out in a different industry. The fastest-growing LED manufacturers weren’t the GE’s and Philips’s of the world. They were the pure-play LED companies that didn’t have legacy fluorescent and HID product lines to protect. The incumbents eventually caught up, but the pure-plays captured the early margin premium. The same dynamic is playing out in GPU cloud infrastructure.
CoreWeave is the most concentrated bet on GPU scarcity you can make in the public markets. Last week’s Q4 earnings sharpened both the bull and bear cases.
The headline numbers: $5.13 billion in 2025 revenue, up 168% year-over-year, making CoreWeave, in CEO Mike Intrator’s words, “the fastest cloud in history to reach $5 billion in annual revenue.” Q4 revenue hit $1.57 billion, up 110% YoY, slightly beating consensus. Revenue backlog swelled to $66.8 billion, up $11.2 billion sequentially and more than $50 billion year-over-year. The weighted average contract length has extended to five years, up from four at the end of 2024. NVIDIA’s $2 billion strategic investment closed in January, and contracted power expanded to approximately 3.1 gigawatts, with over 850 MW already active. Intrator said the company doubled its reserved instance customer count in Q4 versus any prior quarter and delivered more than 50,000 Grace Blackwells to affected customers after quickly clearing the data center delays disclosed in Q3.
The forward guidance is where the debate starts. Management guided FY2026 revenue to $12-13 billion, roughly in line with the $12.09 billion analyst consensus but below the $2.29 billion Q1 estimate ($1.9-2.0 billion guided). More pointedly, CoreWeave projected $30-35 billion in 2026 capex. Intrator told CNBC he was “willing to take a short-term margin hit” because “clients are desperate to get access to more infrastructure faster.” Margins will start at low single digits in Q1, expanding sequentially to low double digits by Q4, with management reiterating confidence in 25-30% adjusted EBITDA margins over the long term as mature contracts generate mid-20s contribution margins. They also flagged an upcoming revenue stream from licensing CoreWeave’s proprietary cloud stack to other NVIDIA cloud, enterprise, and sovereign customers (not yet in guidance, but potentially a software-margin kicker).
The backlog composition tells the story: OpenAI, Meta, and NVIDIA represent the dominant share of committed revenue. Recent deals with Poolside (40,000+ GPUs, estimated at $5-6 billion over five years), Runway, and now Perplexity AI add incremental diversification. Analyst consensus projects an anticipated annualized run-rate of $17-19 billion exiting 2026, growing to over $30 billion by end of 2027.
CoreWeave’s co-founder previously disclosed that the company signed a multi-year A100 renewal (for GPUs introduced in 2021) at roughly 95% of the original contract price. That’s a powerful data point for GPU useful-life durability and underscores just how supply-constrained the market remains.
But the bear case is equally important to understand, and Q4 brought it into sharper focus. The word is leverage.
CoreWeave’s debt-to-equity ratio now stands at approximately 894%, nearly double the 485% figure from just a quarter earlier. The company posted a $452 million net loss in Q4 (vs. $51 million a year ago) and a $1.2 billion net loss for full-year 2025. Diluted loss per share of $0.89 missed the $0.21 consensus significantly. Free cash flow was negative $4.6 billion for the trailing twelve months. The $30-35 billion 2026 capex guide implies persistent negative free cash flow and continued reliance on external financing. Management raised $18 billion in debt and equity during 2025, reduced the weighted average interest rate by 300 basis points, increased the revolving credit facility to $2.5 billion, and says no near-term debt matures before 2029 (only self-amortizing contract-backed and vendor financing). The market reacted harshly: shares fell as much as 21% the day after earnings, and a class action lawsuit was filed within days.
NVIDIA’s involvement mitigates the financing risk substantially. The $2 billion equity investment and 5 GW development commitment represent genuine strategic backing. But there’s a dependency risk. CoreWeave’s model only works as long as NVIDIA views it as a preferred distribution channel. Bernstein initiated coverage this week with an Underperform rating, citing hyperscaler competition risk and calling the stock “meaningfully overvalued.”
Customer concentration adds another layer of risk. Three customers (OpenAI, Meta, and NVIDIA) dominate the backlog. If OpenAI shifts more compute toward Oracle (it has a reported $300 billion commitment) or if Meta brings more infrastructure in-house, CoreWeave’s backlog conversion could face headwinds.
The stock tells the story of this tension. From its $40 IPO in March 2025, shares surged to $187 before falling back hard. At the current price around $75, the stock is up roughly 88% from IPO but sits approximately 60% below its all-time high. Analyst price targets range from $38 (bear) to $251 (bull), with a consensus around $122—roughly 65% upside from current levels. Market cap is approximately $39 billion against an enterprise value of $67 billion.
If CoreWeave is the startup that became an AI infrastructure giant, Oracle is the legacy software company choosing to become one. That choice may be the more interesting investment story.
Think of Oracle as Amazon without the logistics business. Both are building massive cloud infrastructure operations. But Amazon’s retail logistics is a capital-intensive, low-margin drag. Every dollar of AWS margin partially subsidizes fulfillment centers and delivery vans. Oracle’s legacy is the opposite: a high-margin enterprise software annuity that generates roughly $25 billion in annual operating cash flow. That cash machine funds the AI buildout.
The numbers are staggering, and they just got better. Oracle Cloud Infrastructure revenue hit $4.9 billion in F3Q26 (ending February 2026), beating the $4.74 billion estimate and accelerating from $4.1 billion last quarter. Cloud revenue (IaaS + SaaS) reached $8.9 billion, up 44% year-over-year, now representing 52% of total revenue. Total adjusted revenue was $17.19 billion, up 22%, beating the $16.89 billion consensus. Non-GAAP EPS of $1.79 crushed the $1.23 estimate by 46%. Adjusted operating margin came in at 43%, above the 42.7% estimate. This was Oracle’s first quarter in over 15 years where organic total revenue and non-GAAP EPS both grew at 20% or more. Total remaining performance obligations reached $553 billion—up $30 billion sequentially, up 325% year-over-year. Management reaffirmed FY26 revenue at $67 billion and capex of $50 billion, while raising FY27 revenue guidance to $90 billion (up from $89 billion). A number that seemed aggressive when first floated but now looks increasingly conservative given OCI’s 81% constant-currency growth rate. Most of the RPO increase came from large-scale AI contracts where Oracle says it won’t need incremental funding because customers either prepay for GPU purchases or supply the GPUs themselves. The prior five-year OCI trajectory—$18 billion, then $32 billion, $73 billion, $114 billion, and $144 billion. That roadmap remains intact, with most of that revenue already booked in reported RPO.
That five-year OCI trajectory—if achieved—would make Oracle’s cloud infrastructure business alone larger than all of CoreWeave by FY27 and approaching the scale of AWS’s current revenue by FY29.
Management’s tone on the call was telling. They stated plainly: “The demand for cloud computing for AI training and inferencing continues to grow faster than supply.” More importantly, they noted that “some of the largest consumers of AI Cloud capacity have recently strengthened their financial positions quite substantially,” a direct reference to the massive funding rounds closed by OpenAI, xAI, and others in recent months. Translation: the $553 billion backlog isn’t vapor. The customers can pay, and demand still exceeds what Oracle can build. Management said these dynamics “enable Oracle to comfortably meet and likely exceed our revenue growth rate forecast for FY27 and beyond.” That’s not a company hedging its guidance. That’s a company telling you the $90 billion FY27 target is conservative.
There’s a second angle here that’s underappreciated. Oracle disclosed that AI code generation has become efficient enough that they’re restructuring product development teams into “smaller, more agile and productive groups”—building more SaaS applications, for more industries, at lower cost. This is an operating leverage story layered on top of the infrastructure story. OCI grows the top line at 81% constant currency while AI-driven development compresses the cost of expanding the SaaS portfolio. Margins expand on both sides of the business simultaneously. CoreWeave can’t replicate that compounding.
But Oracle faces its own version of the funding gap. The company announced a $45-50 billion financing plan for CY26—a mix of equity and debt, because even with growing cash flow, they can’t self-fund $50 billion in annual capex. Oracle currently carries approximately $135 billion in total debt. The good news: within days of the February announcement, Oracle raised $30 billion through investment-grade bonds and mandatory convertible preferred stock, with a record oversubscribed order book. Management said they don’t expect additional bond issuance in CY26, and the $20 billion ATM equity program remains untapped, preserving flexibility. The 43% operating margin and $1.79 EPS beat confirm Oracle is scaling OCI revenue and holding margins simultaneously—killing the “buying revenue with margin-destructive capex” bear case. And crucially, most new RPO comes from contracts where customers prepay or supply their own GPUs, reducing Oracle’s incremental funding burden.
Free cash flow was negative $24.7 billion on a trailing four-quarter basis through F3Q, driven by $48.3 billion in cumulative capex. But operating cash flow grew 13% to $23.5 billion, confirming the underlying business generates real cash even during peak investment. FCF isn’t expected to turn positive until FY29 as the buildout continues.
The critical difference from CoreWeave: Oracle funds through investment-grade credit markets with no additional bond issuance planned beyond what’s already raised. The ATM equity program introduces dilution risk (the near-term bear case) but it’s fundamentally different from CoreWeave’s 894% debt-to-equity financed by GPU-collateralized private credit.
Oracle also has a class action lawsuit to contend with, but of a different nature than CoreWeave’s. Theirs relates to broader corporate disclosures rather than specific infrastructure execution failures.
The channel data backs the hypothesis. Jefferies’ proprietary survey of 20 Oracle partners (published March 2026) paints a picture of a business accelerating into capacity constraints. 85% of partners hit or exceeded F3Q plans, with 35% beating by more than 6%. 75% reported OCI capacity constraints negatively impacting their practices—a slight uptick from 70% the prior quarter, meaning demand is outrunning Oracle’s ability to build. Pipeline improved for 70% of respondents, and partners expect AI initiatives to contribute roughly 6 points of growth in CY26. Most tellingly, partners cited Oracle’s strength in sovereign cloud and regulated cloud environments—the segment where US export policy may soon force foreign entities to co-invest in US infrastructure.
The backlog quality is improving too. OpenAI’s share of RPO declined from 66% in F1Q26 to 57% in F2Q26, still elevated versus peers (Microsoft at ~45%, Amazon at ~40%), but the diversification trajectory is in the right direction. With RPO now at $553 billion after F3Q, that diversification is happening at scale. Jefferies estimates cloud current RPO covers near-term cloud revenue at a 0.99x coverage ratio—meaning estimates are essentially fully backed by contracted obligations. Even excluding OpenAI entirely, Jefferies models IaaS reaching over $85 billion by FY30 at a 53% CAGR, arguing that “capacity reserved for OAI is somewhat fungible and aggregate GPU-driven demand in the industry will continue to increase.” Their DCF yields a $320 per share base case—more than double the current price.
The bears are already poking at RPO composition, and the question deserves a direct answer. Oracle disclosed that most of the F3Q RPO increase came from contracts where customers either prepay so Oracle can procure GPUs, or the customer buys and supplies the GPUs themselves. The question: if the customer brings the hardware, what exactly is in the $553 billion? RPO captures two distinct contract structures. In prepaid deals, the full contract value, including GPU procurement, flows through RPO. In customer-supplied-GPU deals, the RPO reflects the contracted value of everything else: power, cooling, networking, the OCI orchestration stack, managed services, and SLA guarantees over contract terms that now average five-plus years. Oracle didn’t break out the split, and more disclosure would help. But the margin implication actually favors the bulls: customer-supplied-GPU contracts are higher margin for Oracle because the company avoids hardware capex and obsolescence risk entirely. It’s pure platform and infrastructure revenue. And the customer’s own GPUs are physically sitting in Oracle’s facilities, creating switching costs that go beyond the contract term. The bear case on RPO quality inverts when you think about it through a margin lens rather than a top-line lens.
Throughout my Co-Design Series, I’ve argued that the winners in AI infrastructure will be companies that design hardware, software, and deployment as unified systems, not independent components. Let me apply that framework here.
In my Inference Stack Depth piece, I mapped five layers of inference infrastructure: networking, orchestration, model optimization, model architecture IP, and inference distribution. NVIDIA’s strategy is to own all five layers through acquisitions and internal development: Mellanox for networking, Run:ai for orchestration, and its Israel acquisitions for optimization.
CoreWeave operates primarily at Layer 2: Orchestration. Their fleet lifecycle controller, tensorizer, and bare-metal deployment stack are purpose-built for allocating GPU resources across workloads. They do this extremely well, arguably better than any other cloud provider. But they don’t control the network fabric (Layer 1), they don’t develop model optimization tools (Layer 3), and they have no model architecture IP (Layer 4). CoreWeave is a single-layer specialist with world-class execution at that layer.
Oracle operates across multiple layers simultaneously. OCI provides the compute orchestration (Layer 2), but Oracle’s database technology creates something CoreWeave can never replicate: a data-to-GPU pipeline. Multi-cloud database consumption grew 531% year-over-year in F3Q (817% in Q2) because enterprises are running Oracle databases on OCI GPUs for AI reasoning on private data. Oracle’s AI Data Platform unifies all enterprise data and allows reasoning using the latest AI models while keeping that data private and secure.
This is application-infrastructure co-design. When Larry Ellison talks about bringing all three layers of Oracle’s software stack together—infrastructure, database, and applications—to enable AI models to do multi-step reasoning on private enterprise data, he’s describing exactly the kind of stack integration that creates defensible competitive advantages. Not silicon-level co-design like Vera Rubin, but a software-infrastructure form of the same principle.
The F3Q earnings call sharpened this argument considerably. CEO Clay Magouyrk stated it plainly: enterprises aren’t training their own LLMs. What’s “incredibly popular and growing in popularity is people taking the best models and wanting to combine that in a private way with their private data.” That’s the demand signal, and it changes the competitive framing entirely. This isn’t about who has the fastest GPU cloud. It’s about who owns the layer where foundation models meet proprietary enterprise data. Oracle added MCP server connectivity and natural-language-to-SQL capabilities to its AI Database, making it trivially easy for enterprises to connect the best models to their private data without moving it to a competitor’s cloud. The AI Data Platform product pulls together application data, data lakes, lake houses, and structured databases into “an agentic platform to quickly build applications on, as well as access to all of the greatest models from multiple providers.”
If this sounds familiar, it should. In Is Software Dead?, I wrote: “Context is the battleground. Situational awareness of a problem, understanding of humans involved, business goals, priorities. Systems of record dominate digital context today—your CRM knows your customers, your ERP knows your inventory, your HRIS knows your employees.” And: “Agency—the ability to act autonomously—requires context as a precursor. You can’t delegate a task to an AI agent without giving it the context to execute. This is why systems of record are more valuable in an AI world, not less: they’re the context that makes agency possible.” Oracle’s AI Data Platform—the MCP server connectivity, the natural-language-to-SQL, the agentic platform Magouyrk described on tonight’s call—is that evolution happening in real time. They’re not building a system of record. They’re building a system of context. I also wrote that “SaaS companies must evolve from ‘systems of record’ to ‘systems of context.’ The companies that can provide rich context to AI agents will capture the orchestration layer. Those that can’t will be reduced to dumb data stores.” Oracle owns the richest enterprise context of any cloud provider—and they’re running it at 43% operating margins while most SaaS companies are still figuring out what a token costs them.
Here’s why this matters for the investment thesis: when an enterprise runs AI workloads on AWS, they’re sending their most sensitive data to a company that also operates a retail marketplace, an advertising platform, and a competing custom silicon program. When they run on Google Cloud, they’re trusting the world’s largest advertising company with proprietary data. Oracle has no competing data business. Chip neutrality extends to data neutrality: Oracle doesn’t monetize customer data because it doesn’t have an ads or retail business to feed. For regulated industries—banking, healthcare, government, sovereign institutions, that distinction isn’t a feature. It’s a requirement.
Magouyrk also flagged the migration tailwind: “For customers to take advantage of the latest and greatest AI, they first have to be in the cloud, and there’s still a lot of data that’s not in the cloud.” The 531% MultiCloud database growth is the on-ramp—enterprises moving private data into cloud environments specifically so they can run AI reasoning against it. Oracle is building private token factories for the world’s largest institutions, and the data already lives in Oracle databases. That installed base is the moat no hyperscaler can replicate.
CoreWeave sells raw compute. Oracle sells compute plus a data platform plus enterprise applications as an integrated system. The companies that own multiple stack layers capture more value over time. That’s the pattern I keep seeing across this entire buildout.
To understand why NeoClouds matter, you have to understand why the hyperscalers are structurally slower at pure NVIDIA GPU deployment, even though they’re spending far more money. The spending is extraordinary. The four largest hyperscalers have guided to approximately $710 billion in combined 2026 capex: Amazon at $200 billion, Alphabet at $175-185 billion, Microsoft at roughly $150 billion, and Meta at $115-135 billion. Add Oracle’s $50 billion, and you’re approaching $760 billion in a single year.
But here’s the critical question: how much of that goes to NVIDIA GPUs versus custom silicon, logistics, real estate, networking, and other priorities?
Amazon (AWS) announced $200 billion in 2026 capex, $50 billion above expectations. The stock dropped 11%, erasing $450 billion in market value. Amazon doesn’t disclose the split between NVIDIA GPUs, custom silicon, and non-AI infrastructure, but the signals are mixed. AWS’s custom chip business reportedly reached a $10 billion annual run rate with triple-digit growth. Trainium2 has over a million chips deployed, and Project Rainier placed 500,000 Trainium2 chips for Anthropic. Trainium3 is now generally available on TSMC’s 3nm, and Trainium4 was announced with NVLink Fusion support, meaning AWS is building hybrid clusters where its own silicon plugs into NVIDIA’s interconnect. But here’s what’s telling: Amazon’s press releases for major customer wins consistently lead with NVIDIA GPUs, not Trainium. The NVLink Fusion integration is itself an acknowledgment that customers want the NVIDIA ecosystem. Trainium appears to be growing as an additive option for specific workloads, not a wholesale replacement for NVIDIA. And the logistics business, fulfillment centers, delivery networks, and Project Kuiper satellites all compete for that $200 billion.
Google (GCP) has made the most aggressive custom silicon move of any hyperscaler, and the story has changed dramatically since most investors last looked. Google launched TPU v7 Ironwood in late 2025, delivering 4,614 teraflops FP8 per chip with 192 GB of HBM3e. A single Ironwood pod connects 9,216 chips delivering 42.5 exaflops—scaling to hundreds of thousands of chips. Anthropic committed to up to 1 million TPUs in a deal worth tens of billions, with 400,000 chips purchased outright through Broadcom and 600,000 rented through GCP in a deal estimated at $42 billion in RPO. Meta is reportedly in advanced discussions for a multi-billion-dollar TPU lease starting 2026. According to SemiAnalysis, Google’s TPU v7 delivers approximately 44% lower total cost of ownership than GB200 for Google, and even after adding Google’s margin, external customers see roughly 30% lower TCO than equivalent NVIDIA infrastructure. Google Cloud backlog hit $240 billion (up 55% sequentially, doubled year-over-year), and the company guided $175-185 billion in 2026 capex, double its 2025 spend. Cloud revenue is growing 48% on a $70 billion+ run rate. The two best frontier models—Claude and Gemini 3—now run the majority of their training and inference on Google TPUs and Amazon Trainium. This is a structural shift, not a side project.
Microsoft (Azure) just launched Maia 200 in January 2026, its second-generation custom AI accelerator built on TSMC’s 3nm process with 140+ billion transistors. The chip delivers over 10 petaflops FP4 and 5+ petaflops FP8 with 216 GB of HBM3e at 7 TB/s bandwidth. Microsoft claims it’s the “most performant first-party silicon from any hyperscaler” with 3x the FP4 performance of Trainium3 and FP8 performance above Google’s TPU v7. Maia 200 is already deployed running GPT-5.2 and powering Microsoft 365 Copilot. But the path was messy: the chip was delayed six months due to design changes requested by OpenAI that caused instability in simulations, and as many as one-fifth of staff on some chip design teams left. The earlier Maia 100 reportedly never powered any production AI services and was used internally only for staff training. On top of the custom silicon complexity, Azure must balance capacity across the OpenAI partnership commitments, millions of enterprise customers, and the Copilot ecosystem consuming inference capacity internally.
None of this means the hyperscalers will lose. They have capital advantages, customer relationships, and geographic reach that NeoClouds can’t match for years. But it does mean they’re playing a fundamentally different game: a diversified, multi-front war across three or four competing silicon ecosystems simultaneously (NVIDIA GPUs, custom ASICs, and general-purpose CPUs), not a concentrated bet on NVIDIA GPU deployment speed.
The numbers tell the story. Of Amazon’s $200 billion capex, a meaningful and growing share goes to Trainium and non-GPU infrastructure. Of Google’s $175-185 billion, roughly 60% goes to servers (split between TPUs and GPUs) and 40% to data centers and networking. Microsoft’s Maia program consumes engineering bandwidth and data center floor space. Meta is building custom silicon while simultaneously deploying the largest external NVIDIA GPU clusters.
NeoClouds don’t have this problem. When NVIDIA ships Rubin, CoreWeave and Oracle don’t have to decide whether it competes with their own silicon programs for floor space. They just deploy it.
For readers who follow my work, the “So What?” always connects technical analysis to investment implications. Here’s how I think about the NeoCloud thesis:
If you believe GPU scarcity persists through 2027-2028—which my Memory Wall thesis and Packaging Paradox analysis strongly support—then both NeoClouds benefit. Every CoWoS bottleneck, every HBM supply constraint, every power availability limitation reinforces the value of companies that can deploy available GPUs fastest. CoreWeave offers asymmetric upside from a smaller base; Oracle offers a larger, more diversified expression of the same thesis.
If you believe enterprise AI adoption requires integrated data + compute platforms—which the Co-Design framework argues, then Oracle has structural advantages CoreWeave can’t replicate. The database-to-GPU pipeline, the multi-cloud interoperability, the enterprise sales force: these create compounding lock-in that pure GPU leasing doesn’t.
If you’re worried about the custom silicon threat, the counterargument is important. Google’s Ironwood, Amazon’s Trainium3, and Microsoft’s Maia 200 are real—they’re shipping in volume and winning production workloads. The two best frontier models in the world (Claude and Gemini 3) run primarily on TPUs and Trainium, not NVIDIA GPUs. If custom silicon adoption accelerates to the point where NVIDIA GPU scarcity eases, NeoClouds lose their premium. This is the bear case, and it’s more credible today than it was six months ago. But the NeoCloud hypothesis doesn’t require NVIDIA to have zero competition—it requires NVIDIA to remain the default for the broadest set of AI workloads. CUDA’s ecosystem moat, the breadth of model support, and the frictionless deployment experience mean that for most enterprises and AI labs, NVIDIA remains the path of least resistance. NeoClouds capture that demand.
If you’re worried about financing risk, the distinction matters enormously. CoreWeave’s 894% debt-to-equity ratio and $30-35 billion 2026 capex guide versus Oracle’s investment-grade access represent fundamentally different risk profiles for what is economically a similar activity: borrowing to build GPU data centers. Oracle’s equity dilution is a headwind, but it’s a manageable headwind from a $438 billion market cap base. CoreWeave’s leverage is a potential binary event.
If you think about this through cycles, the question is what happens when GPU supply eventually catches up to demand. CoreWeave’s business model depends on scarcity—when GPUs are abundant, the premium for bare-metal access compresses. Oracle’s business model has a floor: even if GPU margins normalize, the enterprise database and applications businesses generate substantial recurring revenue. As I wrote in Is Software Dead?: “This is why systems of record are more valuable in an AI world, not less: they’re the context that makes agency possible.” Oracle owns that context layer for thousands of enterprises. One is a cycle play; the other is a structural compounder with cyclical upside.
My own framework suggests the market is underpricing Oracle’s NeoCloud transformation and overweighting CoreWeave’s execution risk. Oracle trades at approximately 13x EV/CY27E EBIT versus Microsoft at 16x, despite now compounding both organic revenue and non-GAAP EPS at over 20%. Jefferies maintains a $320 PT based on 20x FY29E EPS of $16. AI infrastructure gross margins came in at 32% in F3Q, within the 30-40% guided range, with higher-margin adjacent services and database mix improving the blend. CoreWeave at ~$75 with 894% debt-to-equity and $30-35 billion in 2026 capex is the higher-beta expression of the same macro thesis, with 65% upside to consensus. Appropriate only for investors who size for the leverage embedded in the name.
Both benefit from the same tailwind: when NVIDIA ships Rubin, these two install it first. The question is which risk profile matches your portfolio.
The NeoCloud category is a direct consequence of the co-design era I’ve been writing about. As NVIDIA’s platform becomes more integrated—silicon, software, networking, and orchestration as a single system, the value of being NVIDIA’s preferred deployment partner increases. The companies that can take a Vera Rubin NVL576 rack and have it serving inference workloads within weeks of delivery capture disproportionate value in the early months of each GPU generation.
This is reminiscent of what happened in my previous industry. When Cree released a new generation of LED chips, the lighting manufacturers that could integrate them into products fastest captured the premium. The slow-movers eventually caught up, but by then the fast-movers had the customer relationships and the margins.
The AI infrastructure buildout is the same pattern at 1000x the scale. NeoClouds are NVIDIA’s fastest channel to market. Whether you express that thesis through CoreWeave’s concentrated leverage or Oracle’s diversified platform, you’re betting on the same fundamental dynamic: in a supply-constrained environment, speed of deployment is the scarce resource.
And as I wrote in The Verification Gap: the great unbundling of the cloud era is giving way to the great re-bundling of AI infrastructure. The competition to own that bundle has only begun.
There’s a new wrinkle that could accelerate the NeoCloud thesis further. The US Commerce Department is reportedly considering a new four-tier framework for AI chip exports, where shipments above 200,000 units would potentially require companies to invest in US AI data centers or meet stricter security conditions. If this takes shape—and the regulatory momentum suggests it will—it creates a structural tailwind for US-based NeoClouds. Sovereign AI buyers from the Middle East, Southeast Asia, and beyond who want large chip allocations would need to co-invest in US-based infrastructure. Oracle, with its 147 sovereign cloud regions and its role in the Stargate JV, is purpose-built for this. CoreWeave, as the pure NVIDIA deployment partner, becomes the natural landing pad for GPU-intensive sovereign commitments. Even the Jefferies partner survey flagged it: “Oracle is strong in Sovereign Cloud / regulated Cloud environments.” If US export policy effectively channels foreign AI infrastructure investment through domestic NeoClouds, the demand tailwind becomes structural.
BEP Research: NVIDIA’s Inference Stack Depth Strategy (Co-Design Series Part 2)
BEP Research: The Verification Gap (Co-Design Series Part 3)
If you found this analysis valuable, please share it. This is the last free piece before GTC coverage begins next week.
If you tried to subscribe recently and your payment was declined (especially Apple Card or Apple Pay), that’s been fixed. It was a payment processor error on our end. I’m extending the early-bird discount through end of day Friday, March 13th.
Subscribe to BEP Research (discount extended through Friday) →
GTC is next week and I’m locked in. I’ll be interviewing Adel El Hallak (VP, Product Management, Agentic AI for the Enterprise) and Dion Harris (Sr. Director, HPC and AI Infrastructure Solutions) at NVIDIA. Expect some great content coming out of that.
Thanks for the patience. — Ben
Ben Pouladian is a Los Angeles-based tech investor and entrepreneur focused on AI infrastructure, semiconductors, and the power systems enabling the next generation of compute. He was co-founder of Deco Lighting (2005–2019), where he helped build one of the leading commercial LED lighting manufacturers in North America. Ben holds an electrical engineering degree from UC San Diego, where he worked in Professor Fainman’s ultrafast nanoscale optics lab on silicon photonics and micro-ring resonators, and interned at Cymer, the company that manufactures the EUV light sources for ASML’s lithography systems.
He currently serves as Chairman of the Leadership Board at Terasaki Institute for Biomedical Innovation and is a YPO member. His investment research focuses on AI datacenter infrastructure, GPU computing, and the semiconductor supply chain. Long-term NVIDIA investor since 2016.
Follow on Twitter/X: @benitoz | More at benpouladian.com
Disclosure: The author holds long ORCL 2027 LEAPS and positions in NVIDIA and related semiconductor investments. The author does not hold a position in CoreWeave (CRWV). This is not investment advice. Do your own research.
Originally published on BEP Research on Substack. Subscribe for more.
Leave a Reply