NVIDIA Q2 FY2026 snapshot (actual world)

1. Total revenue: ~$46.7B (+56% YoY)
2. Data Center: $41.1B (~88%)
3. Gaming: ~$4.3B (~9%)
4. Pro Visualization: ~$0.6B (~1.3%)
5. Automotive & Robotics: ~$0.6B (~1.3%) NVIDIA NewsroomTom's HardwareInvestopedia

That mix tells the story: modern NVIDIA is overwhelmingly a Data Center company.

The counterfactual: NVIDIA never enters Data Center

Imagine Jensen keeps NVIDIA focused on Gaming + Automotive—no A100/H100/Blackwell, no NVLink/NVSwitch/NVL racks, no CUDA-first cloud boom. What changes?

1) The platform moat shifts: from CUDA to open compilers and cloud-native silicon

Without CUDA dominating training workloads, the software stack converges earlier around OpenXLA/MLIR/IREE and vendor-agnostic backends. PyTorch and JAX would have pushed equal-first support beyond CUDA much sooner, making portability a default—not an afterthought. Outcome: less vendor lock-in, faster heterogeneity.

2) “King of AI” becomes a three-headed crown

a) Google (TPU) — the default for frontier training

TPU becomes the de facto training substrate for large models. With no NVIDIA alternative at hyperscale, research and foundation model labs consolidate on TPU v4/v5e/v5p, and PyTorch-TPU support matures rapidly. Google’s vertical integration (chips + compiler + data center + model tooling) puts it in pole position for both internal and external workloads.

b) AMD — the merchant-silicon champion

Freed from CUDA gravity, ROCm/HIP gains earlier parity on frameworks and kernels. MI200/MI300 hit scale faster with cloud distribution, and AMD’s tight coupling to HBM supply (and TSMC advanced nodes) lets it become the go-to open hardware alternative for both training and inference. Expect stronger early wins with Meta, Microsoft, Oracle, and sovereign AI builds.

c) Cloud custom silicon — the default for inference (and eventually training)

AWS Trainium/Inferentia, Google TPU, Microsoft Maia/Cobalt, and Meta MTIA absorb the lion’s share of inference, then move up-stack into large-scale training for in-house models. With no NVIDIA, hyperscalers’ TCO and supply chain imperatives accelerate their in-house roadmaps 1–2 years.

3) Networking & systems stack: Broadcom, Arista, and Ultra Ethernet rise faster

No NVLink/NVSwitch dominance means Ethernet-first AI fabrics win early. Broadcom (Tomahawk/Trident) and Arista consolidate switching share; Ultra Ethernet Consortium-style enhancements reach production sooner. The “in-node” bandwidth gap (where NVLink excelled) is addressed by tighter HBM + package co-design and compiler-level parallelism.

4) Specialists expand the frontier

With the CUDA path absent, specialized accelerators gain more oxygen:

1. Google TPU: frontier training standard outside merchant ecosystems.
2. Cerebras (wafer-scale): long-context and sparse training niches.
3. Groq (LPU): ultra-low-latency inference and determinism for agentic apps.
4. Tenstorrent/SambaNova: domain-specific training/inference pockets.

5) Economics: AI capex spreads wider, earlier

Hyperscaler spend still explodes, but it splits: TPU and cloud chips dominate training; AMD and cloud chips split inference; specialists capture high-value niches. HBM suppliers (SK hynix, Samsung, Micron) remain the real winners. TSMC still leads advanced packaging and 3D integration, but merchant share tilts more toward AMD and cloud chips.

6) What happens to NVIDIA in this timeline?

1. Gaming remains a cultural and revenue pillar—but consumer cycles cap growth.
2. Automotive/Robotics/Edge (Jetson-class) scale meaningfully, yet nowhere near data center economics.
3. Omniverse/Simulation finds industrial traction, but without the data center flywheel, growth is steadier, not parabolic.
4. Bottom line: a strong company—but not the gravitational center of AI.

Who’s the “king of AI” in the no-NVIDIA-DC world?

Crown split, roles distinct:

1. Google/TPU: Frontier training king (vertically integrated, research-first).
2. AMD: Merchant-silicon champion (open stack, broad ecosystem, price/perf leader).
3. Cloud custom silicon: Inference emperor (TCO, scale, control), progressively encroaching on training.

The AI era still arrives—just more pluralistic, more open, and with fewer single-vendor moats. Innovation speeds up in compilers and interconnects; supply risk is lower; switching costs go down. The trade-off? Slightly slower early momentum (fewer drop-in parts like H100/NVL72), but a healthier, more competitive long-run landscape.

Reality check (back to today)

In our actual timeline, NVIDIA’s data center bet concentrated the stack—accelerating time-to-scaled AI while creating a colossal moat. Q2 FY2026 shows how dominant that choice became: ~88% of revenue from Data Center on a $46.7B quarter. That concentration is precisely what would have been diffused across TPU, AMD, and cloud chips in the alternate world. NVIDIA NewsroomTom's Hardware

Sources

1. NVIDIA Q2 FY2026 press release and IR pages (revenue and segment figures). NVIDIA NewsroomNVIDIA Investor Relations
2. Tom’s Hardware roundup on Q2 results (segment detail and 88% data center share). Tom's Hardware
3. Investopedia Q2 coverage (segment breakdown confirmations).