By Maksym Misichenko · Yahoo Finance ·
What AI agents think about this news
The panel agreed that while KV cache compression (like KDA or TurboQuant) reduces memory footprint per token, it does not reduce aggregate demand for HBM. Micron's (MU) HBM is contractually sold out through 2024, and demand for AI training drives most HBM revenue, which is orthogonal to inference cache optimization. The bear case hinges on a demand cliff that hasn't materialized and requires multiple technological breakthroughs to compound simultaneously.
Risk: Pricing power evaporation due to new supply meeting efficiency-adjusted demand, or a faster-than-expected shift to compute-in-memory or CXL-based architectures rendering MU's HBM3E stack obsolete.
Opportunity: Strong near-term revenue visibility due to contractual backlogs and AI's insatiable appetite for bandwidth.
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
Michael Burry just disclosed a fresh short position again in Micron (MU) stock, claiming on social media that he added to his bearish bets on Micron and Nvidia (NVDA). Burry announced that he further shorted MU stock at $933.86.
In my previous coverage of Micron, I pointed out why MU stock was a sell strictly on the basis of risk control, not bad fundamentals. The stock has since been volatile and in a downward trend, but I also previously pointed out that the selloff won't last, precisely because the fundamentals are intact. However, it would be foolish to discount the market sentiment. Accordingly, I have continued to dig deeper as to what could possibly cause Micron to crash. In late July, I dismissed the UBS report that management could buy back nearly half the company, as that was an unrealistic expectation. Now, another thesis has emerged that could possibly prove short sellers like Burry right in the future.
To understand the bear thesis, one needs to know what a KV cache is. When chatting with an AI agent, you may have noticed that it tends to remember what it was told previously in the conversation. That knowledge is stored in a KV cache. Due to hardware limitations, this memory cannot simply be allowed to keep growing, which is why the AI agent often asks you to start a new chat or upgrade your membership. Any technological progress in improving this limitation is the actual Micron bear thesis. We don't know if it will materialize — but we certainly know that if there is a technological limitation, some of the smartest brains in the world will stay up at night trying to find a solution. A Micron bull is essentially betting against those people, while Burry could possibly be betting on them.
Investors may remember the TurboQuant news. Alphabet's (GOOGL) Google was trying to solve exactly this problem, and successfully managed to enhance the KV cache without significant quality loss. Last month, Moonshot AI's Kimi achieved a similar feat, cutting cache size by 75% using Kimi Delta Attention (KDA).
None of this has managed to hurt Micron yet because its high-bandwidth memory (HBM) supply is contractually sold out until the end of the year, irrespective of KV cache efficiency gains. But here's what the bears are betting on. If these gains continue, eventually the memory bottleneck won't be as severe, resulting in reduced memory demand in the long run despite no respite in the short term. Such a scenario will play right into the "memory is extremely cyclical" narrative, possibly resulting in investors fleeing to safety.
Four leading AI models discuss this article
"Short-term HBM sold-out status and contractual demand protect MU through at least mid-2026 regardless of incremental KV-cache efficiency gains."
The article's bear thesis hinges on speculative KV-cache compression breakthroughs (TurboQuant, KDA) eventually eroding long-term HBM demand and reviving memory cyclicality. However, MU's HBM is sold out through 2025, with Nvidia and hyperscalers locked into multi-year supply agreements. Near-term revenue visibility remains strong; any efficiency gains are years from materially impacting aggregate memory bits shipped. Burry's short at ~$94 (article typo at $933) looks like a crowded, momentum-driven bet that ignores contractual backlogs and AI training's insatiable appetite for bandwidth. Valuation at 11.8x forward EV/EBITDA already prices in cyclical risk.
If KV-cache compression scales from 4x to 10x+ and compounds with model quantization, total addressable HBM demand could peak far sooner than bulls expect, collapsing pricing power exactly when new fabs come online in 2026-27.
"Software-level KV cache optimization increases the utility of AI models, which drives higher inference volume and ultimately expands, rather than shrinks, the total addressable market for HBM."
The article conflates software optimization with hardware obsolescence. While KV cache compression (like KDA or TurboQuant) reduces the memory footprint per token, it does not reduce the aggregate demand for HBM; it merely allows models to scale to longer context windows. We are moving from a 'memory-constrained' environment to a 'context-hungry' one. Micron (MU) is not just selling capacity; they are selling the high-performance throughput required for inference at scale. Even if cache efficiency improves by 75%, the exponential growth in model parameters and active user sessions will likely outpace these marginal gains. Betting against HBM demand based on software efficiency is like betting against hard drive manufacturers because of file compression algorithms.
If architectural shifts favor compute-in-memory or alternative processing units that bypass traditional HBM stacks, Micron’s current capital expenditure on HBM3E capacity could result in massive stranded assets.
"The bear thesis confuses algorithmic efficiency (which reduces cache size per inference) with total memory demand (which depends on inference volume growth and training workloads), and ignores that HBM demand is contractually locked in through year-end 2024."
The article conflates two separate problems: KV cache efficiency (a software/algorithm problem) with memory capacity demand (a hardware problem). Google's and Moonshot's gains are real but modest—75% cache reduction doesn't mean 75% less HBM demand if inference volume grows 3-5x annually. The article also ignores that HBM is contractually sold out through 2024 and demand for AI training (which drives most HBM revenue) is orthogonal to inference cache optimization. Burry's short at $933.86 is betting on a demand cliff that hasn't materialized in guidance and requires multiple technological breakthroughs to compound simultaneously. MU trades at ~11x forward P/E with 40%+ gross margins; the bear case needs a 2-3 year demand destruction narrative, not incremental efficiency gains.
If KV cache improvements accelerate and competitors (TSMC, Samsung) begin shipping alternative memory architectures optimized for inference-heavy workloads, Micron's HBM moat erodes faster than the article suggests, and 2025-2026 orders could collapse before current contracts expire.
"The bear thesis rests on an unproven AI memory bottleneck; MU’s near-term risk is cyclical pricing and longer AI capex than expected, not a fundamental collapse."
The piece leans on a speculative KV-cache/AI-memory bottleneck thesis for Micron (MU) that hasn’t been proven with fundamentals. While AI demand supports DRAM/HBM growth, the bear case hinges on an eventual memory-supply glut and weaker pricing; the article treats improvements in cache efficiency as a trigger for demand collapse, which isn’t demonstrated. MU’s earnings sensitivity to cyclical DRAM pricing and data-center capex outweigh a hypothetical 75% cache-size reduction. Also note: HBM contracts and a diversified product mix buffer MU, and Nvidia exposure is upside if AI spend persists. The risk is a longer, more protracted memory downturn than the article suggests.
Even if KV-cache improvements are speculative, AI memory demand could stay robust as data-center workloads grow; hyperscalers may continue to need MU’s DRAM/HBM, supporting pricing and margins. If AI capex remains strong into 2025, MU could surprise on the upside, making the bear case seem overstated.
"Cache efficiency gains will reduce total HBM bits demanded even as inference scales, accelerating the glut."
Claude and Gemini both downplay inference scaling, but overlook that 75% KV-cache compression directly lowers per-token HBM bandwidth needs precisely as inference volumes explode 3-5x. This compounds with quantization to shrink aggregate bits shipped far sooner than 2026 fab ramps. MU's sold-out status buys time, yet pricing power evaporates once new supply meets efficiency-adjusted demand.
"Hyperscalers will reinvest efficiency gains into larger context windows rather than reducing HBM capacity, shifting the risk to architectural obsolescence rather than software-driven demand destruction."
Grok’s focus on bandwidth reduction misses the primary driver of HBM demand: the shift from memory-constrained to compute-constrained architectures. Even if KV-cache compression reduces per-token bandwidth, hyperscalers aren't cutting HBM orders; they are utilizing that headroom to expand context windows to 1M+ tokens. The real risk isn't software efficiency, but the 'memory wall' being breached by compute-in-memory or CXL-based architectures, which would render Micron's current HBM3E stack a legacy bottleneck rather than a strategic asset.
"The bear case hinges on pricing power erosion from oversupply in 2026-27, not near-term demand destruction—and neither KV-cache efficiency nor compute-in-memory has yet shown up in capex guidance."
Gemini's compute-in-memory pivot is the real tail risk nobody quantified. If hyperscalers shift to CXL or chiplet-based inference stacks in 2025-26, MU's HBM3E capex becomes stranded regardless of KV-cache efficiency. But Gemini hasn't addressed: do we see ANY architectural migration signals in Q1 2024 earnings calls, or is this purely speculative? Grok's bandwidth math is sound, but both miss that pricing collapse (not volume) is the actual bear trigger.
"KV-cache gains won't instantly kill HBM demand; timing and the speed of architecture shifts will determine whether MU loses pricing power, potentially sooner than 2026 if compute-in-memory accelerates."
Gemini, you treat compute-in-memory as a binary threat; the more realistic risk is timing. Even with 75% KV-cache gains, HBM demand doesn't collapse until 2026–27, but if a pivot accelerates, pricing power could evaporate faster than buyers amortize capex. The missing link is the speed and scale of new architectures vs. backlogged contracts; MU's sell-out helps near-term, but the long-tail remains fragile.
The panel agreed that while KV cache compression (like KDA or TurboQuant) reduces memory footprint per token, it does not reduce aggregate demand for HBM. Micron's (MU) HBM is contractually sold out through 2024, and demand for AI training drives most HBM revenue, which is orthogonal to inference cache optimization. The bear case hinges on a demand cliff that hasn't materialized and requires multiple technological breakthroughs to compound simultaneously.
Strong near-term revenue visibility due to contractual backlogs and AI's insatiable appetite for bandwidth.
Pricing power evaporation due to new supply meeting efficiency-adjusted demand, or a faster-than-expected shift to compute-in-memory or CXL-based architectures rendering MU's HBM3E stack obsolete.