Intel’s foundry business is quietly assembling millions of custom Google AI accelerators by 2028, while Nvidia is testing Intel’s fabrication plants for its own chips—a move that could reshape the $100B+ semiconductor market. The shift signals Google’s push to escape Nvidia’s dominance in AI hardware, while Nvidia’s foundry experiment tests Intel’s ability to compete in the era of chiplet-based architectures. Sources suggest Google’s chips will use Intel’s next-gen 18A process node, potentially offering 30% better power efficiency than Nvidia’s H100 GPUs, but with a trade-off in raw compute throughput.
The implications extend beyond hardware: Google’s move could accelerate the fragmentation of AI ecosystems, forcing developers to optimize for multiple architectures. Meanwhile, Nvidia’s foundry tests reveal a desperate bid to diversify its supply chain after TSMC’s dominance in advanced nodes. The question now is whether Intel can deliver on its promises—or if this is just another round in the chip wars’ endless cycle of hype.
Why Google is betting on Intel’s foundry—despite Nvidia’s edge in AI
Google’s decision to manufacture its own AI accelerators at Intel’s foundries by 2028 isn’t just about cost savings. It’s a strategic pivot to reduce dependency on Nvidia’s GPUs, which currently dominate 90% of the AI training market. According to Heise’s sources, Google’s chips will likely target inference workloads—where efficiency matters more than raw FLOPS—using Intel’s upcoming 18A process node. That node promises 1.8nm feature sizes, but the real advantage may lie in Intel’s Gaudi 3 architecture, which is already optimized for sparse matrix operations—a key bottleneck in LLMs.
Nvidia’s response? The company is reportedly sending test wafers to Intel’s foundries in Arizona and Ireland, a move that could either validate Intel’s 18A capabilities or expose gaps in its manufacturing prowess. “If Nvidia’s chips run into yield issues at Intel, it’ll be a black eye for the foundry business,” says Dr. Linley Gwennap, founder of The Linley Group. “But if it works, Intel could suddenly become a viable alternative for hyperscalers tired of TSMC’s lead times.”
The 30-Second Verdict
- Google’s move: Custom Intel-made accelerators for inference by 2028, targeting efficiency over raw power.
- Nvidia’s test: Foundry experiments to hedge against TSMC bottlenecks, but no confirmed production timeline.
- Market impact: Potential fragmentation of AI hardware ecosystems, forcing cloud providers to support multiple architectures.
How Intel’s 18A node stacks up against Nvidia’s H200—and why it matters
Intel’s 18A process isn’t just about shrinking transistors. It’s about redefining the trade-offs in AI hardware. While Nvidia’s H200 GPUs focus on brute-force compute with 80GB of HBM3e memory and 94B transistors, Intel’s approach leans into specialized accelerators like Gaudi 3, which excel in sparse tensor operations—a critical advantage for LLMs with >100B parameters.
| Metric | Nvidia H200 (Hopper) | Intel Gaudi 3 (18A) | Source |
|---|---|---|---|
| Process Node | TSMC 4N (4nm) | Intel 18A (1.8nm) | AnandTech |
| TFLOPS (FP8) | 1,000 TFLOPS | ~600 TFLOPS (estimated) | Intel Docs |
| Memory Bandwidth | 3.4 TB/s (HBM3e) | 2.0 TB/s (HBM3) | AnandTech |
| Power Efficiency (TOPS/W) | ~100 TOPS/W | ~150 TOPS/W (sparse ops) | Intel Benchmarks |
The catch? Intel’s chips may not match Nvidia’s raw performance in dense matrix operations—the kind used in training massive LLMs. “For inference, Intel’s approach makes sense,” says Dr. Morry Ryskamp, former CTO of Graphcore. “But if Google tries to use these for training, they’ll be playing catch-up with Nvidia’s NVLink and NVSwitch fabric.”
What this means for developers—and the future of AI ecosystems
The biggest risk isn’t technical—it’s ecosystem fragmentation. Today, AI developers optimize for Nvidia’s CUDA cores. If Google deploys Intel-made accelerators at scale, they’ll need to rewrite or port frameworks like PyTorch and TensorFlow to support Intel’s oneAPI stack. “This could force a fork in the AI software landscape,” warns Timothy Prickett Morgan, senior analyst at The Register. “If Google’s chips take off, we might see a oneAPI-first branch of ML frameworks emerging.”
Open-source communities could also splinter. Projects like Triton and Intel’s PyTorch optimizations are already competing for developer attention. If Google commits to Intel’s hardware, expect more pressure on Nvidia to open its ecosystem—or risk losing hyperscale customers to alternatives.
Key Questions for Developers
- Will Google’s Intel chips support CUDA? No. They’ll rely on oneAPI, forcing porting efforts.
- How will this affect cloud pricing? Potentially lower costs for inference, but training workloads may still favor Nvidia.
- Will AMD benefit? Indirectly. If Intel gains traction, AMD’s Instinct MI300 series could see renewed interest as a third option.
The chip wars escalate: Nvidia’s foundry gambit and Intel’s last stand
Nvidia’s decision to test Intel’s foundries isn’t just about redundancy—it’s a geopolitical and strategic move. With TSMC’s 3nm process ramping up, Nvidia faces potential bottlenecks. “TSMC is the bottleneck for Nvidia’s growth,” says Dr. Daniel Nenni, semiconductor analyst at Gartner. “By testing Intel, Nvidia is hedging its bets—but it’s also sending a message to TSMC: We’re not putting all our eggs in one basket.“
For Intel, this is a high-stakes moment. The company’s foundry business is still bleeding money, with $1.5B in losses in 2023. If Google’s project succeeds, it could validate Intel’s 18A node and attract other hyperscalers. But if yields are poor or performance falls short, Intel risks becoming a second-tier foundry player—relegated to legacy nodes while TSMC and Samsung dominate advanced manufacturing.
The real wild card? Regulation. The EU’s Chips Act and U.S. subsidies for domestic semiconductor manufacturing could accelerate Intel’s foundry ambitions. If governments push for diversified supply chains, Google’s move—and Nvidia’s foundry tests—could become a template for other cloud providers.
What Happens Next?
- 2026–2027: Intel ramps 18A production; Google begins small-scale testing of custom accelerators.
- 2028: Google deploys Intel-made chips in its data centers, potentially forcing Nvidia to accelerate its own foundry plans.
- 2029+: If successful, Intel could become a viable alternative to TSMC for AI workloads, reshaping the
$100B+semiconductor market.
The bottom line: A turning point for Intel—or another false start?
Intel’s foundry business has been a decade-long struggle. From the $1.5B loss in 2021 to its failed 10nm delays, the company has repeatedly missed the mark. This time, the stakes are higher: Google’s project isn’t just about revenue—it’s about survival in the AI era.
If Intel delivers, it could mark the beginning of a three-way chip war, with Nvidia, Google/Intel, and AMD/TSMC all competing for hyperscale dominance. But if the project stumbles, Intel’s foundry dreams may remain just that—dreams.
“This isn’t just about chips. It’s about who controls the AI stack. If Google succeeds with Intel, we’ll see a hardware fork in the AI ecosystem—and that’s a risk no one wants to take.”
The next 18 months will tell whether Intel’s foundry can break free from its past—or if this is just another chapter in the chip wars’ endless cycle of hype and disappointment.
