Rivals gain ground in inference and custom silicon, yet an innovation void is hard to support
Alternatives to Nvidia AI chips are real and growing, yet none can replace Nvidia in training within two years. Jensen Huang’s scandal created a negotiating opening and left the technology gap unchanged.
Huang resigned on 12 November. Buyers, investors and rivals asked whether another supplier could step in. TF takes the view of an industry analyst, a futurist and a neutral observer. The analysis separates a leadership crisis from a chip roadmap.
What’s Happening & Why It Matters
Is There an Innovation Void?
The phrase assumes Nvidia stopped innovating, and the evidence for a stall is thin. Nvidia’s board expects operations and customer commitments to continue without interruption, per Yahoo Finance. Huang’s exit is a governance crisis, which TF covered in its Huang report.
The real gap is leverage. Large buyers have a fresh reason to ask for a second source. Joe Kaiser of Switchyard Partners, quoted by Fortune, compares the moment to Intel after Andy Grove stepped down in 1998. Kaiser expects Nvidia to keep AI training for two or three years, given CUDA switching costs.
How the Market Read the News
On 12 November, AMD rose 5.79% to $584.89, Broadcom gained 5.65%, and Alphabet gained 2.06%, per Yahoo Finance. Nvidia fell 4.59% the same afternoon. Yahoo adds the gains do not show Nvidia lost customers, because the alternatives existed before Huang resigned.
Share prices measure expectations, and shipments tell a different story. One AOL Finance opinion piece says Microsoft, Meta, Amazon and Alphabet supply 40% to 50% of Nvidia’s revenue. All four build custom chips. The piece gives unverified figures, so treat the range as a signal.
Merchant Rivals: AMD and the Specialists
AMD has the strongest merchant case. Forkast reports OpenAI and Meta reconfirmed a combined 12 GW commitment at AMD’s July event. The first 1 GW of MI450-class hardware is due in the second half of 2026. The Helios rack holds 72 MI455X GPUs and 31 TB of pooled HBM4 memory.
AMD still faces open questions. Forkast flags the maturity of the ROCm 7 software stack against CUDA. Forkast names dependence on TSMC’s 2nm and 3nm ramps and on HBM4 supply as risks.
Fortune names Cerebras, Etched and Qualcomm as specialist options. Tom’s Hardware says Huawei’s Ascend 910C targets 600,000 units in 2026 with yields near 20%. Intel cut Gaudi 3 shipment goals by more than 30% in 2024.
Custom Silicon: Google, Amazon and Broadcom
Custom chips hold the largest share of the challenge. Tom’s Hardware cites projections of 27.8% ASIC-based AI server shipments in 2026. ASIC shipments grow 44.6%, against 16.1% for merchant GPUs. Broadcom and Marvell control about 95% of custom ASIC co-design.
Google says Ironwood TPUs cost about 44% less to run than a GB200 server, a vendor claim. Anthropic committed to up to one million TPUs, per the same report. Amazon Web Services reports deploying more than one million Trainium chips, and Trainium4 will support Nvidia’s NVLink Fusion. Broadcom has a $73 billion AI backlog (about €65 billion). OpenAI targets a first custom-accelerator deployment in the second half of 2026.
Internal chips come with limits. Meta says its MTIA chips complement Nvidia, and Meta expanded its Nvidia deal in February 2026. Microsoft claims Maia 200 delivers 30% better performance per dollar. Maia 100 never ran production AI at scale, and Maia 200 slipped about six months.
Three Walls Rivals Must Climb
Software is the first wall. CUDA has a long head start, and each rival needs developers to port code with little pain. Capacity is the second. Tom’s Hardware estimates Nvidia holds about 60% of TSMC’s advanced packaging allocation, against 15% for Broadcom and 11% for AMD.
Proof is the third wall. Most speed and cost figures come from the vendors, and buyers need independent tests. Inference gives rivals the best opening, since Deloitte projects it will account for two-thirds of AI compute in 2026. TF’s single provider analysis explains why buyers seek a second source.
My take: Rivals will not fill an innovation void, because none exists. They will fill a bargaining gap, and inference is the first place they win it. By 2030, I expect a split market. Nvidia leads training and premium systems, and custom chips carry most inference at the large clouds. Buyers gain the most, since every credible rival bid improves the price of the next Nvidia order.
TF Summary: What’s Next
Nvidia reports earnings on 17 November, and the call will show whether customers changed orders. AMD’s first MI450 shipments, Broadcom’s OpenAI deployment and Trainium4 all arrive between late 2026 and the first months of 2027. Independent benchmarks will decide which vendor claims hold.
MY FORECAST: Custom ASICs will exceed 30% of AI server shipments in 2027, up from a projected 27.8% in 2026. Nvidia will still lead AI training share on 31 December 2028.
Related Stories
- Jensen Huang Resigns as Nvidia CEO After 33 Years
- TF Analysis: The Hidden Cost of Single Provider Dependency
- Nvidia Stock Falls as Board Opens Probe Into CEO Jensen Huang
- Gemini 4 Hydrogen Replaces Gemini 3.8 Flash as Google’s Everyday AI Model
- Microsoft 365 Outage Hits Outlook and Teams Worldwide