Let me look at Dell's quarterly revenue pattern and the April quarter specifically to understand what drove that near-doubling.I can see the quarterly pattern but I need to understand what happened around the EMC merger. Let me also look at the FY17–FY18 transition more carefully and check the Dell/EMC deal.The data makes the story clear. Here's what happened:

Dell closed the $67 billion acquisition of EMC Corporation on September 7, 2016 — the largest tech merger in history at the time. EMC brought with it a federation of businesses including VMware, Pivotal, RSA, and Virtustream.

The revenue impact is stark in the quarterly data:

The near-doubling from ~$51B (FY16) to ~$62B in FY17 (which only included ~5 months of EMC), and then to $79B in FY18 (first full year with EMC) is entirely structural — this wasn't organic growth. Legacy Dell pre-merger was a ~$50–55B PC/server company; EMC added roughly $25B in annual revenue overnight, mostly from storage, virtualization (VMware), and data center infrastructure. The jump is simply consolidation math, not a surge in demand.

You're right, let me focus on the right period. Let me look at what drove the revenue jump in the most recent TTM period.Now let me check what drove the April quarter spike and get segment-level detail.Now let me also get the Q2 FY26 (July quarter) results to complete the picture.Now I have the full picture. Here's the real story:

Revenue didn't "almost double" — looking at your annual table, revenue went from $95.6B (Jan-25) to $113.5B (Jan-26), about a 19% increase. But TTM revenue through the April 2026 quarter tells a more dramatic story because of the timing of the AI server inflection.

The step-change happened in Dell's fiscal Q2 2026 (the July 2025 quarter). Quarterly revenue jumped from ~$23.4B to $29.8B — a single-quarter record and +19% YoY. Here's why:

AI-optimized servers exploded. In Q2 FY26, Dell's Infrastructure Solutions Group (ISG) hit $16.8B in revenue, up 44% YoY, driven by record AI server shipments. Dell shipped roughly $7B in AI servers in that one quarter alone. To put that in context:

• Dell booked $12.1B in AI server orders in Q1 alone — more than it shipped in all of FY25
• Exited Q1 with a $14.4B AI server backlog
• The July quarter (Q2) was when those orders began converting to revenue at scale
• Full-year FY26 guidance was raised to $105–109B, with $15B+ in AI server shipments expected

The key customers: hyperscalers and AI startups like xAI (Elon Musk), CoreWeave, plus expanding enterprise adoption across financial services, manufacturing, and sovereign AI (including the U.S. Department of Energy's NURSE-10 supercomputer).

So the "almost doubling" effect you're seeing in TTM is the mechanical result of layering these massive AI server quarters ($29.8B in Jul-25, $27.0B in Oct-25) on top of the prior year's much smaller quarters (~$22–24B range). The traditional PC/server business (CSG) grew only low-to-mid single digits — AI infrastructure is virtually the entire growth story.

That's a sharp observation, and you're right — the hyperscalers do build their own. But the dynamic is more nuanced:

They design, but they don't fully self-manufacture. Google, Amazon, and Microsoft design their own motherboards, racks, power delivery, and networking fabrics — but they don't fabricate silicon or do mass assembly in-house. What they buy from Dell (and Supermicro, and Quanta) is largely white-box integration at hyperscale: Dell assembles to the hyperscaler's spec, leveraging its supply chain muscle, manufacturing footprint, and — crucially — access to NVIDIA GPUs.

The real bottleneck is GPU allocation. NVIDIA doesn't sell H100s and Blackwells to anyone with a credit card. Allocation is political and relationship-driven. Dell, as the #1 server vendor by volume, gets priority access to GPU trays. Even hyperscalers that build their own racks often source GPU-integrated nodes through OEMs like Dell because:

1. NVIDIA prefers selling full reference designs (HGX, DGX) through OEM partners rather than piecemeal to hundreds of end customers
2. Dell handles the messy integration — liquid cooling, power distribution, NVLink fabric, firmware validation, BIOS — at a scale few can match
3. Enterprise and Tier-2 cloud players (CoreWeave, Lambda, xAI, Oracle Cloud) lack the internal manufacturing capability entirely and buy turnkey

The hyperscaler split: Google builds much in-house. Microsoft and Amazon use a mix — some self-build, some OEM. But the marginal AI capacity being added right now is coming disproportionately from Tier-2/sovereign/enterprise buyers who buy fully integrated Dell boxes. And even Tier-1 hyperscalers find that during an allocation crunch, going through Dell gets them GPUs faster than trying to go direct.

So Dell's AI server revenue is less about selling to Google's data center team and more about selling to everyone else racing to build AI capacity — and occasionally to hyperscalers themselves when speed matters more than margin.

Good list — but it's worth separating who actually buys servers versus who consumes compute.

Who buys the servers (Dell's actual customers):

• xAI (Elon Musk) — Confirmed Dell customer. Built its Memphis Colossus cluster with Dell AI servers. Musk has publicly acknowledged this.
• CoreWeave — GPU-specialist cloud provider, one of the biggest Dell AI server buyers. They rent NVIDIA GPUs to AI labs and enterprises.
• Lambda Labs — Similar model to CoreWeave. GPU cloud, buys Dell.
• Oracle Cloud (OCI) — Uses Dell for portions of its AI infrastructure buildout.
• Microsoft Azure — Mixture. Builds some in-house, but also buys OEM through Dell/Supermicro when speed demands it, especially during the initial GPT-4/Copilot infrastructure scramble.
• Sovereign AI / Government — U.S. DoE (NURSE-10 supercomputer confirmed Dell), plus national AI cloud builds in Singapore, Japan, Middle East, Europe.
• Enterprise directly — Large banks, manufacturers, pharma companies buying on-prem AI clusters rather than renting cloud.

Who doesn't typically buy Dell servers:

• OpenAI — Doesn't buy metal. They run on Microsoft Azure, which provisions capacity. Azure may or may not use Dell underneath, but OpenAI never touches the hardware procurement.
• Anthropic — Same story. Runs on AWS (primarily) and Google Cloud. They consume instances, not servers.
• Together AI, Fireworks, Groq — Inference/API providers. Most rent GPU capacity from CoreWeave, Lambda, or hyperscalers. They don't buy iron directly — that's the whole point of their business model (they abstract away infra).
• Google DeepMind / Google Cloud — Google builds almost everything in-house (TPUs, custom servers).

The pattern: Dell's AI customer list skews toward infrastructure providers, not model builders. The GPU cloud operators (CoreWeave, Lambda) and the vertically-integrated AI companies (xAI) buy servers. The model labs (OpenAI, Anthropic) and inference APIs (Together, Fireworks) consume compute as a service and never see a rack.

This is the existential question hanging over the inference-as-a-service space. The honest answer: for vanilla open-weight model serving, the moat is thin and getting thinner. But there are real differentiators for the ones executing well:

Where the defensibility is weak:

• "We run vLLM on rented GPUs" — Anyone can do this. Together, Fireworks, Anyscale, and a dozen others all started here. If that's still the pitch in 2026, they're in trouble.
• Open-weight model serving (Llama, Mistral, etc.) — Commodity. AWS Bedrock, Azure AI, and GCP Vertex all offer the same models with better latency (co-located with enterprise data) and no egress fees. Hard to beat the hyperscalers at their own game when you're renting their own GPUs.

Where some actually have a moat:

1. Inference optimization at the kernel level. Fireworks, in particular, rewrites CUDA kernels and builds custom serving stacks (not just vLLM wrappers). They can compress models, fuse operations, and squeeze 2–3x more throughput per GPU. That margin delta is everything when you're renting by the hour.

2. Multi-model routing / compounding. Some are building layers above raw inference — chaining models, caching semantic embeddings, routing queries to the cheapest model that can handle the task. Groq does this with deterministic hardware-software co-design. This starts to look less like "hosting" and more like an inference platform.

3. Fine-tuning and custom model deployment at scale. Together AI's real business increasingly isn't just inference — it's training and fine-tuning as a service, with LoRA adapters deployed at scale. The inference becomes sticky when a customer has 100 custom fine-tuned models deployed.

4. Latency edge. GPUs rented from CoreWeave in the same datacenter as the inference provider get lower latency than a customer going to CoreWeave themselves, because the inference provider colocated the hot model weights. Small edge, but real for chat/agent use cases.

5. Hardware-accelerated inference. Groq and Cerebras aren't renting NVIDIA GPUs at all. They built their own silicon. That's genuinely defensible — Groq's LPUs deliver token generation speeds that NVIDIA hardware literally cannot match, because they architected for inference rather than training. Cerebras similarly delivers 1,800 tokens/second on Llama 70B, which is absurd. Nobody can replicate that on rented H100s.

The likely outcome: The "me too" inference APIs (rent GPUs + vLLM + serve Llama) consolidate or die. The survivors will either own silicon (Groq, Cerebras), own the full training-to-inference pipeline (Together), or get acquired by hyperscalers for the kernel-level optimization talent. It's a classic infrastructure layer squeeze — the hyperscalers commoditize the middle, and only the hardware-advantaged or deeply integrated survive.