AI & Software
AI & Software Insights
Notas de investigación sobre resultados, estructura de mercado y estados financieros. Cada pieza empieza con la conclusión y luego presenta la evidencia y las implicaciones.
2026-07-15
%22%2F%3E%0A%20%20%3Ccircle%20cx%3D%22220%22%20cy%3D%22150%22%20r%3D%22180%22%20fill%3D%22%2322d3ee%22%20fill-opacity%3D%220.12%22%2F%3E%0A%20%20%3Ccircle%20cx%3D%221010%22%20cy%3D%22130%22%20r%3D%22220%22%20fill%3D%22%2338bdf8%22%20fill-opacity%3D%220.08%22%2F%3E%0A%20%20%3Ccircle%20cx%3D%22920%22%20cy%3D%22520%22%20r%3D%22170%22%20fill%3D%22%2322d3ee%22%20fill-opacity%3D%220.06%22%2F%3E%0A%20%20%3Cpath%20d%3D%22M140%20430h920%22%20stroke%3D%22url(%23line)%22%20stroke-width%3D%2212%22%20stroke-linecap%3D%22round%22%2F%3E%0A%20%20%3Cpath%20d%3D%22M220%20360c120-150%20260-220%20380-220s260%2070%20380%20220%22%20fill%3D%22none%22%20stroke%3D%22%2367e8f9%22%20stroke-opacity%3D%220.5%22%20stroke-width%3D%2216%22%20stroke-linecap%3D%22round%22%20stroke-linejoin%3D%22round%22%2F%3E%0A%20%20%3Cpath%20d%3D%22M300%20398h600%22%20stroke%3D%22%2394a3b8%22%20stroke-opacity%3D%220.24%22%20stroke-width%3D%228%22%20stroke-linecap%3D%22round%22%2F%3E%0A%20%20%3Ctext%20x%3D%22220%22%20y%3D%22510%22%20fill%3D%22%23e2e8f0%22%20fill-opacity%3D%220.7%22%20font-family%3D%22Arial%2C%20Helvetica%2C%20sans-serif%22%20font-size%3D%2238%22%20letter-spacing%3D%228%22%3EINSIGHTS%3C%2Ftext%3E%0A%3C%2Fsvg%3E)
GPU / NPU / ASIC / TPU / CPU en Profundidad: Cómo Nvidia, AMD, Broadcom, Alphabet, Apple e Intel Están Reconfigurando el Mercado de Cómputo de IA de $400B - Rendimiento, Cuellos de Botella, Cadena de Suministro y Demanda
El mercado de cómputo de IA en 2026 es un mercado de $400B+ segmentado en 5 arquitecturas vinculantes: GPU (Nvidia H100/H200/B100/B200/GB300 + AMD MI300/MI325/MI350), NPU (el acelerador de IA on-device más limpio para el Apple Neural Engine + Qualcomm Hexagon + Intel NPU), ASIC (Broadcom ASIC personalizado para hiperescaladores + Marvell silicio personalizado), TPU (Alphabet TPU v5/v6/v7) y CPU (Intel Xeon + AMD EPYC + Apple M-series). Esto es una inmersión profunda full-stack: benchmarks de rendimiento (throughput FP8 / FP16 / BF16 / INT8 / FP4), la cadena de suministro HBM + CoWoS + SiP, los principales expertos, la base de clientes, el capex y la lectura hacia adelante para el ciclo de capex de IA.
HBM Deep Dive: How SK hynix, Samsung Electronics, and Micron Built the Single Most Concentrated AI Supply Chain of the Cycle - and Why the 2026-2028 Capacity Is Already Sold Out
HBM (High-Bandwidth Memory) is the binding AI accelerator supply chain constraint, and the 2026-2028 supply is already spoken for. SK hynix leads with ~52% HBM market share, Samsung Electronics is at ~20%, and Micron is at ~28% with the cleanest 2026 ramp. This is a full-stack deep-dive into HBM: TSV technology, stack height (HBM2/2E/3/3E/4/4E), 12-Hi/16-Hi stacks, base die logic process, the HBM3E race, capacity sold out through 2026, top experts at SK hynix's M16, expansion capex, the customer concentration (Nvidia, AMD, Broadcom), and the read-through for the 2026-2028 AI capex cycle.
Liquid Cooling Deep Dive: How Vertiv, Boyd, Rittal, CoolIT Systems, and Asetek Built the Most Concentrated AI Data Center Thermal Stack of the Cycle
Liquid cooling is the binding AI data center thermal stack of the cycle - air cooling is no longer sufficient for AI accelerators >700W TDP, and the 2026-2028 liquid cooling demand is the cleanest single read on the AI capex cycle's thermal leg. Vertiv leads with ~30% liquid cooling share, Boyd at ~10%, Rittal at ~10%, CoolIT Systems (the cleanest pure-play, recently acquired by Eaton) at ~8%, and Asetek at ~5%. This is a full-stack deep-dive into liquid cooling: cold plate (DLC) vs immersion (single-phase + two-phase) vs rear-door heat exchanger (RDHx), the coolant chemistry (PG-25, dielectric fluids), the customer base (Microsoft, Alphabet, Amazon, Meta), the top experts, the capex, the 2026-2028 supply-demand, and the read-through for the AI capex stack.
NAND Flash Deep Dive: How Samsung Electronics, SK hynix, Kioxia, Sandisk, and Micron Run the Only Memory Market That Stopped Behaving Like a Commodity
NAND flash is the only memory sub-segment where AI demand did not just lift ASPs - it broke the commodity cycle itself. Sandisk and Micron reported datacenter-grade NAND revenue inflection in 2026, Samsung Electronics and SK hynix are running near-full-utilization, and Kioxia is the latest target of a $15-20B IPO/pre-IPO secondary at a ~$20B mark. This is a full-stack deep-dive into the NAND industry: cell architecture (SLC/MLC/TLC/QLC/PLC), node roadmap, capital intensity, controller + firmware stack, top experts, expansion difficulty, debt load, customer concentration, and the read-through for Apple, Nvidia, the AI capex stack, and the 2026-2030 storage market.
Optical Module Deep Dive: How Coherent, Lumentum, Innolight, Eoptolink, and Source Photonics Run the $40B 800G / 1.6T Optical Module Cycle - and Why 3.2T CPO Is the 2027-2028 Binding Test
Optical module is a $40B/year market in 2025 that is structurally being reshaped by the 800G / 1.6T cycle and the emerging 3.2T co-packaged optics (CPO) cycle. The 5 leading-edge optical module suppliers are Coherent, Lumentum, Innolight, Eoptolink, and Source Photonics. This is a full-stack deep-dive: 800G / 1.6T / 3.2T optical module + EML + DML + silicon photonics (SiPh) + CPO + linear pluggable optics (LPO) + tunable laser + EDR / FR / DR / SR / LR / ER / ZR reach classes, the EML + SiPh + CPO technology stack, the Nvidia + Broadcom + Cisco + hyperscaler customer base, the 2025-2028 capex, the top experts, and the read-through for the AI capex cycle.
PCB Deep Dive: How Nvidia AI Server Boards, Unimicron + Shennan Circuits Substrate-Like-PCB, and the 800G / 1.6T Switch Era Are Reshaping the $80B Global PCB Market
PCB (Printed Circuit Board) is a $80B/year global market in 2025 that has been structurally reshaped by the Nvidia AI server board cycle. The 5 leading-edge PCB suppliers are Unimicron, Shennan Circuits (the cleanest China PCB exposure), Tripod Technology, Shinko Electric, and Ibiden. This is a full-stack deep-dive: substrate-like-PCB (SLP) + high-density interconnect (HDI) + any-layer HDI (aHDI) + modified semi-additive process (mSAP), the 800G / 1.6T switch PCB cycle, the 224 Gbps PAM4 signaling, the Nvidia GB200 NVL72 / GB300 / Rubin class boards, the customer base (Foxconn + Quanta + Wistron + Inventec + Wiwynn), the capex, and the read-through for the AI server cycle.
SK Hynix's 10% Single-Day Crash Made Korea the First Asia Tape to Reprice the Micron Memory Sell-Off
SK Hynix shares plunged over 9% in Seoul on July 16, 2026, reversing the prior session's 8% rally, as the Micron -8% sell-off in the U.S. overnight spilled into Asia. The Kospi dropped 6.78% to 6,790.23, with Samsung Electronics -7%, Seoul Semiconductor -5%, and Tokyo Electron -5%. The Bank of Korea simultaneously hiked rates 25bp to 2.75% - the first hike since January 2023 - against a backdrop of June CPI at a three-year high of 3.2% and a won near 1,484. The combined tape is the cleanest stress test of the AI-memory trade outside the U.S. since the 2024 cycle peak.
Stripe and Advent's $53B Bid for PayPal Is the First Mega-Deal That Says Fintech M&A Is Back
Stripe and Advent International made a $53 billion all-cash takeover offer for PayPal at $60.50 per share on July 15, 2026, sending PYPL up 17.20% to close at $55.52 on 72.5 million shares - roughly 7x its 10-day average volume. The bid is the first mega-deal of the post-SpaceX IPO window and re-prices the entire fintech M&A funnel: Block, Adyen, Robinhood, Toast, Payoneer, and Western Union all move with this trade. The bid premium, the consortium structure, and the unanswered response from PayPal's board all signal that fintech consolidation is no longer theoretical.
Tesla acaba de limitar el gasto en IA de los ingenieros a 200 $/semana; luego, Musk les ordenó usar su propio modelo Grok, que ocupa el puesto 9
Electrek informó el 10 de julio de 2026 que Tesla implementó un tope semanal de 200 $ para el gasto en IA de terceros: notablemente eximiendo a Grok de xAI; y, después, Elon Musk envió un memorando en el que instruía a los empleados de Tesla a cambiarse a Grok, el modelo de IA de su empresa xAI (ahora fusionada en SpaceX). Grok 4.5 ocupa el puesto 9 en los rankings multiámbito con una puntuación de codificación de 68,6 (la más baja entre los modelos comparados), por detrás de OpenAI (GPT-5.6 Sol/5.5/5.6 Terra/5.4), Anthropic (Claude Fable 5, 4,8 Opus, 4,7 Opus) y Google (Gemini 3.1 Pro). Se informa que los ingenieros de Tesla prefieren a Claude de Anthropic para el trabajo diario de desarrollo. La lectura es directa para Tesla, Anthropic, OpenAI, Alphabet, Meta y todo el grupo empresarial de IA que ahora navega preocupaciones por desvío de negocios.
Trump vs Hochul's NY Data-Center Moratorium Is the First Real Test of Whether AI Capex Has a Political Ceiling
President Trump publicly slammed New York Governor Kathy Hochul's new executive order barring for up to a year the construction of data centers that use 50+ MW of power, the first state-level moratorium of its kind in the U.S. Trump wrote on Truth Social that New York 'has made a terrible decision' and urged the state to change policy 'IMMEDIATELY.' The moratorium is the first concrete political constraint on the AI capex stack from a U.S. state, and the read-through is direct for Microsoft, Alphabet, Amazon, Meta, Nvidia, Equinix, Digital Realty, and every AI-infrastructure issuer with a NY exposure.
TSMC Pledges Additional $100B for Arizona on Top of Existing $165B - Q2 2026 Net Profit Jumps 77% YoY - The Largest Single CHIPS-Era Capex Announcement in History
CNBC reported on July 15, 2026 that TSMC announced an additional $100B Arizona investment on top of its existing $165B US commitment, on the back of Q2 2026 net profit surging 77% YoY. The new incremental US commitment brings TSMC's total US investment to $265B across three Arizona fabs, making it the largest single CHIPS-era capex announcement in history. The Q2 2026 net profit beat consensus by ~10%, and the $100B incremental capex signals 3-5 year demand visibility for AI training + inference. The read-through is direct for Nvidia, AMD, Apple, Broadcom, Qualcomm, Intel, Applied Materials, Lam Research, KLA, ASML, Samsung Electronics, SK hynix, and the entire US chip sovereignty thesis.
Warsh Just Turned the Fed From a Single-Mandate Inflation Fighter Into a Multi-Task Force Monetary Experiment
CNBC reported on July 15, 2026 that Fed Chair Kevin Warsh used two days of congressional testimony to launch a multi-pronged review of how the Fed measures inflation, restore money-supply reporting, and convene an all-bullish AI task force - all while CPI and PPI both printed unexpected declines. Fed Governor Lisa Cook pointed to AI spending as a potential inflation driver and warned 'inflation risks now outweighing employment risks.' The Fed meets in two weeks and markets are split on whether rates rise, hold, or fall - and the read-through is direct for every rate-sensitive sector, every AI capex issuer, and the bond market that is now repricing the Fed governance premium.
2026-07-14
CleanSpark's $6.6 Billion Lease Turns a Bitcoin Miner Into a Power-and-Compute Landlord
CleanSpark's July 14 SEC filing disclosed a 20-year data-center lease covering 175 megawatts of critical IT load, approximately $6.6 billion of initial-term rent, and about $330 million of average annual net operating income. The market's 8.7% midday response is not just an AI enthusiasm trade: it is a vote that contracted rent from investment-grade compute demand can deserve a very different multiple from volatile bitcoin-mining cash flow.
Hundreds of Economists Are Telling the Market What the AI Trade Does Not Want to Hear
AP reported on July 13 that hundreds of economists signed a joint letter calling for immediate action on AI's economic impact and job-displacement risks. The market is still pricing AI as a productivity story. The economists are forcing a parallel conversation about labor-market disruption, antitrust, and AI-specific regulation that can change the political risk premium on the entire AI trade.
Nvidia's Asia Buyer Crackdown Turns Export Compliance Into a Moat
Nvidia is halving its list of authorized Asian buyers and tightening checks across Singapore, Malaysia, and Japan. The immediate effect is less channel freedom, but the deeper read is that compliance is becoming part of Nvidia's pricing power and its geopolitical firewall.
AI Agents Are About to Kill the SaaS Seat - and Sierra Just Made the Outcome-Pricing Bet Public
CNBC reported on July 14, 2026 that Sierra co-founder Clay Bavor said AI agents are moving from demos into real business workflows - customer service, sales, and support - and that outcome-based pricing could challenge the way software companies get paid. Sierra is the cleanest publicly-discussed example of the AI-agent outcome-pricing model, and the read-through is direct for Salesforce, ServiceNow, Snowflake, MongoDB, Datadog, and the entire enterprise-software seat-based cohort that has been the dominant pricing model for two decades.
2026-07-13
Massive AI Buildout Is Turning Into an Inflation Test Before It Becomes a Productivity Story
The AI spending boom is no longer just a future-growth story. With $720 billion of expected data-center investment, sharply higher memory-chip costs, and visible consumer price increases for laptops and consoles, the market now has to decide whether AI is adding durable productivity or leaking into the inflation data first.
AI Semiconductors Are in a Valuation Reset, Not a Demand Collapse
The July selloff in SK Hynix, Samsung Electronics, Micron, and the broader chip complex looks like a thesis break only if you stop at the tape. The better read is a crowded AI trade digesting geopolitics, index concentration, and expectations that had run ahead of near-term proof.
AI Data Centers Are Hitting the U.S. Grid Before They Hit Scale
Meta's expanded Louisiana build shows hyperscalers are still willing to spend at a shocking scale, but the real bottleneck is now transformers, interconnection queues, and local resistance. The winners are shifting from pure compute landlords to whoever can finance electricity.
OpenAI's Leadership Turnover Shows the Commercial Engine Is Still Not Settled
Fidji Simo's step-back and the latest safety-team departure show OpenAI is still rearranging who owns product, safety, and monetization. That matters because the company is trying to behave like a platform while its executive map keeps changing.
Qué esperar
Notas basadas en evidencia con un punto de vista visible.
Esta sección recoge análisis directos sobre resultados, reuniones de accionistas y estructura de mercado. Cada pieza nueva debe dejar clara la tesis, los hechos y las implicaciones desde las primeras pantallas.
Espera análisis directos, no comentarios genéricos.
Espera datos claros y argumentos fáciles de seguir.