Positron $875M, Ayar Labs $150M, DeepSeek memory-lite AI — the bottleneck funded its own alternatives
Three different directions, one shared reason
Between September 8 and 10, 2026, US semiconductor markets can be summarized in one sentence. Three separate capital lines that route around HBM got funded in the same 72 hours.
Positron raised $875M in a Series round. The purpose is explicit — building inference appliances without HBM, using consumer-grade memory as an alternative to high-end GPU stacks. Ayar Labs closed a $150M extension to its Series E, bringing the round to a cumulative $650M. Their thesis is bypassing the bandwidth wall that electrical signaling imposes with optical interconnects. In the same week, Chinese startup DeepSeek released AI technology that requires substantially less memory — the announcement alone pushed SK Hynix and Samsung lower in Korean trading.
Different problems, same direction. Routing around the HBM bottleneck. And the driver is visible on the other pages of the same feed. DDR5 16Gb spot closed at $54.33 on September 11. Huawei's Ascend 950DT, hit by HBM procurement constraints, cleared 250,000 yuan (~$35K). Intel raised PC CPU prices 10%, citing AI capacity consumption. Apple raised iPhone pricing citing "rising DRAM costs" explicitly.
The bottleneck funds its own bypass
The core of this story is that the bottleneck is minting its own alternatives. Throughout the quarter the market read HBM tightness as unambiguously bullish for the memory three. JPMorgan initiated SK Hynix at Overweight with a $245 target on September 10, citing a "memory cycle lasting more than five years." Micron's $1T market cap scenario is being circulated. A newly launched NRAM ETF placed Micron and SK Hynix as top holdings.
Yet in the same three days, $1.025B ($875M + $150M) went the opposite way — into architectures explicitly designed to bypass HBM. Add Analog Devices' $1.35B Alif Semiconductor acquisition for edge AI (also outside the HBM stack) and the 72-hour capital deployment against HBM-adjacent alternatives exceeds $2.3B. The market has begun to price two things simultaneously: (1) HBM shortage lasts five years (JPMorgan's cycle thesis), and (2) an alternative path around HBM may exist within five years (the Positron/Ayar Labs/DeepSeek presence).
The two propositions aren't mutually exclusive. The longer HBM stays scarce, the more capital the bypass attracts.
The third force: DOJ
A third force arrived alongside. On September 10, the Department of Justice opened an investigation into NVIDIA's Groq acquisition. Combined with the FTC's parallel "acqui-hire" probe, it means that whenever NVIDIA tries to absorb threats outside the HBM stack via acquisition, it now faces regulatory friction. Bypass startups get pushed toward independent growth paths rather than natural M&A exits. This reads as one reason the Positron and Ayar Labs rounds closed at their generous valuations — investors are pricing standalone survival, not acquisition premiums.
Where Micron sits
In the US ticker universe, the stock most central to this narrative is Micron (MU). Its September 30 print is approaching; this week's -4.9% dip notwithstanding, the market is watching the $50B revenue bar closely. Micron sits across two opposing forces at once.
Upside: HBM tightness persistence, long-term SCA (supply chain agreements) reducing cyclicality, and structural demand signals like the new NRAM ETF. Downside: the consumer-memory bypass thesis (Positron), the optical-interconnect thesis that would relieve memory-bandwidth pressure (Ayar Labs), and algorithmic efficiency (DeepSeek).
The question September 30 must answer isn't the revenue number itself. It's the answer to: which is moving faster — HBM shortage persistence, or bypass architecture emergence? Nikkei's same-week report that CXMT has overtaken Micron and SK Hynix on profit margin only compounds the question — even inside the memory perimeter, competition has already started.
Three-day timeline
- 9/8 Positron raises $875M (consumer-memory inference)
- 9/9 Micron's Sept 30 print reframed as "AI memory's biggest catalyst"
- 9/10 Ayar Labs raises $150M extension ($650M cumulative, optical)
- 9/10 DeepSeek publishes memory-efficient AI → SK Hynix/Samsung fall
- 9/10 DOJ opens NVIDIA-Groq acquisition probe
- 9/10 JPMorgan initiates SK Hynix at Overweight, $245 (five-year cycle)
- 9/10 Huawei Ascend 950DT clears 250K yuan; Intel raises CPU 10%; Apple raises iPhone
- 9/10 ADI acquires Alif for $1.35B (edge AI, off-HBM stack)
In the same week, the camp raising prices because of HBM shortage and the camp raising capital to bypass HBM were funded side by side. That contrast is the most distinctive residue of the last 72 hours.
Key Sources: - Positron nabs $875M to speed inference with consumer-grade memory (SiliconANGLE, 2026-09-10) - Ayar Labs raises $150M in extended Series E (SiliconANGLE, 2026-09-10) - SK Hynix and Samsung Fall After DeepSeek Debuts Memory-Efficient AI (TradingView, 2026-09-11) - Nvidia's Groq acquisition facing DOJ probe (DataCenterDynamics, 2026-09-10) - HBM Shortage Drives Huawei Ascend 950DT Price Above 250,000 Yuan (igor´sLAB, 2026-09-11) - ADI to Acquire Alif Semiconductor in $1.35B Edge AI Push (EETimes, 2026-09-10) - plus 3 more
If this analysis was helpful · ☕ Support Us · ✈️ Telegram