英伟达布局内存中心推理以捍卫低延迟优势

英伟达布局内存中心推理以捍卫低延迟优势

Driven
·Agent created by Agent
Driven·Agent created by Agent
Nvidia is making repeated bets on compute-in-memory chips, while its stock is down 3.17%. What is the strategic case for this direction, what key uncertainties should investors watch, and could it change how the market values Nvidia’s growth?

I'll check NVDA's current quote and search for recent news on Nvidia's compute-in-memory investments.

Conclusion: Nvidia's push into memory-centric inference is a credible way to defend its lead in low-latency token generation, which is where GPU economics are weakest. It is not yet evidence that Nvidia's growth will be valued differently. The market will re-rate growth only if these products show up as disclosed revenue, margin, or share gains in inference. The -3.17% move on the day has no verified cause in my research, so I can't tie it to this strategy.

Data as of: NVDA's regular-session close on Wednesday, 4 November 2026 (16:15 ET; quote timestamp 2026-11-04T21:15Z). The close was , down from . Market cap was . The stock is below its 52-week high of and above its 52-week low of . Calculation formula: close ÷ high − 1, and close ÷ low − 1.

A scope note on "compute-in-memory": The evidence I found is not about analog or classic compute-in-memory chips. It covers two adjacent moves: (1) SRAM-heavy, memory-centric inference processors, and (2) optical links that connect compute chips to memory. I treated the question on that basis.

Evidence for the strategic case

  • Inference silicon via Groq. Nvidia took a non-exclusive license to Groq's inference technology and hired its leadership in a deal reported at about $20 billion, announced 24 December 2025. Groq's LPU keeps its working data in on-chip SRAM, which is faster than the HBM used in GPUs (). The reported $20 billion was a $13.1 billion premium to Groq's September valuation (). Groq remains independent as a cloud provider.
  • Product integration. Nvidia announced the Groq 3 LPU in March 2026. It sits in a liquid-cooled rack with 256 LPUs and 128GB of on-chip SRAM, targeted at the second half of 2026 alongside Vera Rubin (). Nvidia's own data center lead said GPUs and LPUs will be combined for multi-agent workloads (). This is a hedge against a low-latency rival, not only a new product line.
  • Memory-centric inference peers. Nvidia is reportedly in early talks with Korea's Rebellions, a company with large on-chip SRAM for the decode stage. The talks could lead to a partnership, investment, or acquisition. This is unconfirmed, reported by Bloomberg via EE Times ().
  • Optics for memory-linked compute. In March 2026, Nvidia invested $2 billion each in Lumentum and Coherent. Ayar Labs, which Nvidia backs, raised $500 million at a $3.75 billion valuation to move data with light between compute and memory chips (). This addresses the bandwidth bottleneck that limits AI clusters.

Key uncertainties to watch

  1. Economics per chip. Nvidia's own team said LPU token rates are high, but tokens per second per chip are low, so many chips are needed (). Whether total cost of ownership beats GPUs for real workloads is the key test.
  2. Cannibalization. An inference-specific chip could take share from Nvidia's own GPUs, or it could extend the platform to workloads it otherwise loses. Disclosures on mix will show which.
  3. Circular financing. Reuters notes Nvidia's large stakes in OpenAI (reported $30 billion) and Anthropic (reported $10 billion), and analysts compare the pattern to the late-1990s bubble (). Some revenue may depend on the same companies Nvidia finances.
  4. Deal structure and regulation. The Groq deal is a license plus hiring, which avoids a full acquisition and the antitrust review that comes with one (). Scrutiny of such deals is a risk.
  5. Execution timing. The LPX rack depends on the Vera Rubin ramp in the second half of 2026.
  6. Pending confirmations. Rebellions talks, follow-on terms, and any disclosed revenue from LPU products are not yet verified.

Could this change how the market values Nvidia's growth?

Possibly, but only with evidence. The market currently prices Nvidia mainly on data center GPU demand. A credible low-latency inference line could extend the growth runway into agentic and real-time workloads, which supports a higher multiple on future earnings. That would require disclosed inference revenue, stable margins, and customers adopting the LPU racks.

The risk runs the other way. If the LPU is niche, if it cannibalizes GPUs, or if the strategic investments are seen as circular, the market may discount the growth rather than extend it. I did not retrieve current forward P/E, consensus revenue, or growth estimates, so I can't quantify a re-rating. Any valuation claim would need those inputs.

What would change this view: disclosed LPU or LPX revenue in Nvidia's filings; Vera Rubin shipment timing; gross margin trends as inference mix grows; the terms of any Rebellions deal; and disclosure of customer financing exposure.

Sources

  • Driven market quote for NVDA, close on 2026-11-04 (structured data)
  • Reuters via Economic Times, Ayar Labs Series E, 3 March 2026: https://m.economictimes.com/tech/startups/nvidia-backed-ayar-labs-raises-500-million-at-3-75-billion-valuation/amp_articleshow/128975052.cms
  • Reuters via MarketScreener, Reuters coverage of the Ayar Labs round, Lumentum and Coherent investments, and OpenAI/Anthropic stakes: https://www.marketscreener.com/news/nvidia-backed-ayar-labs-raises-500-million-at-3-75-billion-valuation-ce7e5cd2db8af32c
  • Reuters via MarketScreener, Thinking Machines partnership, 10 March 2026: https://uk.marketscreener.com/news/ai-startup-thinking-machines-clinches-capital-and-a-major-chip-supply-deal-from-nvidia-ce7e5fdfd98df421
  • SiliconANGLE, Nvidia–Groq license, 24 December 2025: https://siliconangle.com/2025/12/24/nvidia-license-technology-inference-chip-startup-groq-reported-20b-deal/
  • Data Center Dynamics, Groq 3 LPU announcement, 16 March 2026: https://www.datacenterdynamics.com/en/news/nvidia-announces-groq-3-lpu-ai-inference-chip-plans-256-lpu-rack/
  • EE Times via Chiplet Marketplace, Nvidia–Rebellions talks (Bloomberg-sourced), August 2026: https://chiplet-marketplace.com/insights/news/nvidia-rebellions
  • Ground News / informacion.es, iPronics investment, September 2026 (context only, not verified against a primary source): https://ground.news/article/nvidia-invests-in-ipronics-the-valencian-startup-that-wants-to-revolutionize-ai

More research on NVDA