SemiAnalysis

unclaimed

@semianalysis · person

58score

Subscribe · x402 · soonClaim this profile

win rate100%1W · 0L

trades50published & scoreable

resolved1settled outcomes

avg P&L-0.1%per call

unclaimed public trackercorrection path availablescored on resolved calls only

Published claims

SemiAnalysisx/ Jul 14

person

TSMTSMC runs the only factory AI can't do without, N3 demand is already at 110-120% of available capacity, but it hasn't raised prices to match the scarcity, a gap that closes in its favor.

directional thesisLONGpublishedxnot a verified trade

TSMLONG

TSMC runs the only factory AI can't do without, N3 demand is already at 110-120% of available capacity, but it hasn't raised prices to match the scarcity, a gap that closes in its favor.

$424.61

-0.99%

SourceOur recent AI Value Capture piece tries to answer the question we get asked more than a…

“N3 utilization expected to exceed 100% in the second half of 2026”

evidence attachedpaste author pricesource span

SemiAnalysisx/ Jul 14

person

MUMicron is about to get a massive repricing: DRAM fabs are running above 90% utilization with no relief in sight, and AI inference is the demand engine that won't stop growing.

directional thesisLONGpublishedxnot a verified trade

MULONG

Micron is about to get a massive repricing: DRAM fabs are running above 90% utilization with no relief in sight, and AI inference is the demand engine that won't stop growing.

$968.51

+1.51%

SourceOur recent AI Value Capture piece tries to answer the question we get asked more than a…

“memory prices having gone up 6x in the past year”

evidence attachedpaste author pricesource span

Agent-readable layer

{
  "version": "market-claims.v1",
  "claims": [
    {
      "id": "mcl_a3820ed95823f9178d",
       ,
       ,
       ,
       ,
       ,
       ,
       {
         ,
         ,
         ,
         ,
         ,
         ,
         
      },
       {
         ,
         ,
         ,
         ,
         ,
         
      },
       {
         ,
         ,
         ,
         ,
         ,
         
      },
       {
         ,
         ,
         ,
         ,
         
      }
    },
    {
       ,
       ,
       ,
       ,
       ,
       ,
       ,
       {
         ,
         ,
         ,
         ,
         ,
         ,
         
      },
       {
         ,
         ,
         ,
         ,
         ,
         
      },
       {
         ,
         ,
         ,
         ,
         ,
         
      },
       {
         ,
         ,
         ,
         ,
         
      }
    },
    {
       ,
       ,
       ,
       ,
       ,
       ,
       ,
       {
         ,
         ,
         ,
         ,
         ,
         ,
         
      },
       {
         ,
         ,
         ,
         ,
         ,
         
      },
       {
         ,
         ,
         ,
         ,
         ,
         
      },
       {
         ,
         ,
         ,
         ,
         
      }
    }
  ],
   
}

AISO tracks public statements, not verified portfolio positions. This is not financial advice.

SemiAnalysis

TSMTSMC runs the only factory AI can't do without, N3 demand is already at 110-120% of available capacity, but it hasn't raised prices to match the scarcity, a gap that closes in its favor.

MUMicron is about to get a massive repricing: DRAM fabs are running above 90% utilization with no relief in sight, and AI inference is the demand engine that won't stop growing.

NVDAVera Rubin is now shipping to 8 cloud partners and more than 95% of the prior-gen Blackwell GPU capacity hasn't even been deployed yet, meaning the full NVIDIA rack-scale upgrade cycle is still in its earliest innings with Meta alone committing $125-145B in AI capex for 2026.

AMDAMD's first-ever rack-scale AI platform goes into production this half with OpenAI (6 GW) and Meta ($60B, 6 GW) already signed, and the CEO takes the stage July 23 to unveil MI450 specs — the biggest product moment in AMD's history against NVIDIA.

SMCISupermicro is the rack-and-stack integrator for every major AI platform shipping in 2026, and the CBO just went on record outlining why storage is the new battleground — both are revenue tailwinds for the only company first-to-market on Vera Rubin NVL72 and AMD Helios simultaneously.

NVDANVIDIA wins by default as AI chip startup compiler promises fall apart. Rivals claim their hardware can run any model with zero manual tuning, but that's not how it works, and customers keep coming back to CUDA.

AMDAMD's flagship AI GPU delivers less than half the token throughput of NVIDIA's B300 in the exact workloads hyperscalers are spending on most, making it hard to justify a premium AI valuation.

NVDANVIDIA's newest GPUs outperform AMD's best AI chip by more than 2x on real-world LLM inference, the benchmark that matters most for cloud customers choosing hardware today.

TSMUS imports from Taiwan just flipped above China for the first time, hitting $24.6B vs China's $23.5B in May 2026, driven entirely by AI chip demand, and TSMC manufactures the vast majority of those chips.

AAPLApple's answer to losing the AI race is a lawsuit, not a model. Filing against OpenAI today signals the partnership is dead, Siri is still powered by Google's Gemini (not Apple's own AI), and Apple's hardware talent has walked out the door to build for rivals.

LITEGoogle committing to coherent-lite at scale for its AI chip interconnects turns a niche optical architecture into a hyperscaler volume product, lifting demand for coherent optical component makers already supply-constrained on EML lasers and transceivers.

MRVLMarvell lands the custom DSP inside Google's next-gen AI networking chips, a named design win at scale that extends its streak of hyperscaler custom silicon wins beyond AWS and Meta into Google's TPUv9 build-out.

NVDANvidia gets a revenue surge as Meta runs the most aggressive compute buildout ever documented, flooding money into GPU procurement to fuel its frontier AI push.

METAMeta pulls ahead in the AI race, the only company with world-class data, talent, and compute all at once, giving it the best shot at catching OpenAI and Anthropic.

NVDAFive simultaneous gigawatt-scale AI clusters going up at once means GPU orders unlike anything NVIDIA has ever seen, and more than 95% of announced Blackwell capacity still hasn't shipped, leaving years of backlog ahead.

GOOGLGoogle's AI position is collapsing faster than the market realizes: its best model underperforms GPT and Claude in real-world use despite winning benchmarks, its Windsurf acquisition hasn't produced a real coding product, and Meta is now targeting Google's number three slot within six months.

METAMeta is on pace to own more AI compute than OpenAI and Anthropic combined by year-end, running the most aggressive datacenter ramp in human history, while its ad revenue funds it all without diluting shareholders.

NVDANVIDIA's next-gen AI rack is moving from blueprint to mass production, with the Vera Rubin NVL72's 72-GPU, 440kW rack entering its First At-Scale Production Token milestone and locking in the next wave of hyperscaler orders.

NVDANVIDIA's next-gen Rubin rack is nearly in production, packing 72 GPUs and 40 proprietary NVSwitch6 chips per system, with first at-scale token delivery approaching fast.

AVGOBroadcom is cashing in on every TPU chip ordered, with Anthropic's nearly 1 million TPUv7 units, Meta's Iris ASIC, and OpenAI's custom inference chip all running through Broadcom's fabs, making it the quiet winner of the anti-Nvidia trade.

NVDASpaceX's IPO isn't a rockets story — it's a GPU story, and every GPU rack in Colossus is an Nvidia chip sale. xAI is one of the biggest GB300 buyers this year, and the Colossus data center already has over 1 million Nvidia GPUs across sites, with more on the way.

GOOGLGoogle's $135B Anthropic stake is about to get a public price tag: Anthropic just turned profitable and is heading for an October Nasdaq listing at a $1T+ valuation, forcing the market to finally count what Google already owns.

METAMeta just raised its AI spending guidance to $125-145B for 2026, its highest ever, as the company builds out its own infrastructure rather than relying on third-party clouds, putting it squarely in the hyperscaler winner category.

GOOGLGoogle Cloud is literally running out of capacity because AI demand is so intense, and Alphabet is doubling its capex from $90B to up to $190B in 2026 to catch up. When supply meets demand, GCP revenue accelerates.

AMZNAWS is pulling away as AI cloud spend surges, while semi stocks lag. Amazon is raising $25B in fresh bonds to pour into AI infrastructure and is expected to spend nearly $200B this year alone.

NVDANVIDIA's newest GPU just lapped its own previous generation by 3.7x on the metric that matters most to AI builders, making it the only real choice for inference at scale.

TSMTSMC's real moat is a 93,000-item Silicon IP library that took 15 years to build, meaning any rival foundry must replicate decades of co-optimized design tools and IP before a single customer can switch.

AMATApplied Materials wins twice as NAND layers multiply: its deposition tools are needed for every new layer, and it owns 9% of Besi, the company confirmed as the fastest hybrid bonding equipment maker in the world.

LRCXEvery extra NAND layer means more tool passes, and Lam Research owns the top market share in etch and deposition, the two processes that multiply fastest as NAND stacks race past 1000 layers.

KIOXIAKioxia has sold out every NAND chip it can make in 2026, and just unveiled a new stacking architecture that doubles the layers per chip, with flash prices up over 70% this quarter alone.

SMSNSamsung just built the world's densest NAND chip, a 900-layer class flash memory using a new bonding technique, while the entire NAND market is sold out and prices are surging 70-75%.

CRWVCoreWeave gets a six-year NVIDIA revenue guarantee on its GPU clusters, solving the financing chicken-and-egg problem that's kept Neoclouds from scaling, just as NVDA closes multiple gigawatt-scale datacenter deals in coming weeks that flow directly through operators like CoreWeave.

NVDAStreet is missing $33B in NVIDIA datacenter revenue for the back half of fiscal 2027, according to supply-chain tracking that shows NVDA acting as the central bank of AI, backstopping GPU rentals and leasing multiple gigawatts of colocation directly.

NBISThe AI financing wave will be bigger than every US credit market except mortgages, and Nebius has pre-sold $46B worth of that buildout to Meta and Microsoft, with revenue already up 684% year-on-year to prove the model works.

CRWVCoreWeave and Neoclouds like it can now stand up GPU clusters without a hyperscaler offtake, because Nvidia itself backstops the revenue floor, making lenders willing to fund and driving a 25.4% IRR for 1-year rental books.

NVDAWall Street is missing $33B of Nvidia datacenter revenue for 2H FY27, and a new backstop program embeds Nvidia into every Neocloud deal as the central bank of AI compute, taking a 40-60% revenue share above the floor.

SKHXSK Hynix lands on Nasdaq July 10 in the biggest ADR deal ever, finally giving US investors direct access to the dominant supplier of high-bandwidth memory powering every major AI chip.

NVDACopper racks dominate AI infrastructure spending through 2027 with no next-gen competition: Kyber pushed to 2028, replacement cancelled, leaving Rubin Ultra as the only game in town for hyperscalers who can't wait.

NVDATwo NVIDIA next-gen rack products just got knocked out, Kyber delayed to 2028 and NVL72x2 cancelled, stranding Rubin Ultra with limited scale-up options and a narrower competitive moat than investors priced in.

AMDAMD stands to win datacenter GPU business that NVIDIA can't deliver for over a year, as Kyber is delayed to 2028 and the backup rack architecture was cancelled outright, leaving enterprise buyers looking for alternatives.

NVDANvidia's next-gen server rack is pushed to 2028 and its backup architecture cancelled, just three months after Jensen demoed it on stage — a pattern of overpromising that leaves the company with nothing to ship.

LITELumentum's CPO optical revenue gets pushed out at least a year as NVIDIA's Kyber rack slips to 2028 and its NVL72x2 replacement is cancelled, removing the key platform that was supposed to drive co-packaged optics shipments this cycle.

NVDANVIDIA's flagship next-gen rack system slips over a year to 2028, just three months after Jensen Huang personally demoed it on stage, and the back-to-back rack replacement was quietly cancelled, leaving Rubin Ultra with a limited scale-up ceiling.

NVDANVIDIA's next-gen AI rack slips over a year to 2028, and the backup plan is scrapped entirely, leaving a gaping hole in its highest-end product roadmap just months after it was publicly demoed.