Milk Road AI

Jim Cramer strikes again, yesterday he called IBM a buy and this morning it crashed 25%, its worst single day decline since the dot com bubble (Save this). IBM pre announced preliminary Q2 results that came in well below internal expectations. Revenue was $17.2 billion versus $17.86 billion estimated up just 1% year over year. Software grew only 5%, consulting was flat, infrastructure was down 7% and the CEO himself said "this quarter we faltered." The company lost over $65 billion in market cap in a single session. The reason for the miss is what matters for the broader market and it is not an IBM specific story. CEO Arvind Krishna published a letter explaining exactly what happened. "In the last few weeks of June, IBM's enterprise clients abruptly shifted their quarterly capex budgets away from IBM software and mainframe contracts and toward servers, storage, and memory purchases specifically to lock in supply constrained infrastructure ahead of expected price increases. IBM said it "did not anticipate the magnitude of the capex reprioritization." This is the AI capex supercycle cannibalizing enterprise software spending in real time. Companies that were IBM's bread and butter customers for decades are now redirecting discretionary tech budgets toward GPU servers, HBM memory and storage rather than IBM licensing deals and mainframe upgrades. The broader implication is significant. Every enterprise software name, Salesforce, ServiceNow, SAP, Oracle's non-cloud divisions now faces the same risk that IBM just crystallized. When corporate IT budgets are finite and AI infrastructure is the new non negotiable priority, something has to give. Milk Road Pro is tracking every beneficiary of this AI infrastructure wave and every company that's about to get left behind. Come join for just $1 using the link below and position accordingly!

Tom Lee just went on CNBC and said buy every dip in Korea, SK Hynix, and Samsung and here is why he is right (Save this). SK Hynix controls approximately 60% of the global HBM market and reported an operating margin of 72% in Q1 2026 and its entire HBM supply through 2026 is fully contracted. The selloff is technical driven by Korean retail leveraged ETF mechanics not a deterioration in the underlying business. The ADR listing is the structural event that changes the stock's trajectory. SK Hynix listed on Nasdaq under ticker SKHY on July 10, raising $29.65 billion, the largest foreign ADR offering in history, surpassing Alibaba's 2014 debut. Until this listing, most American institutional capital was categorically unable to own SK Hynix. Pension funds, index funds, and insurance companies operating under US listed only mandates had no direct path to the dominant AI memory company in the world and that changes overnight. SK Hynix trades at 6.2x forward P/E versus Micron's 7x, despite holding the superior position in HBM and that discount closes as passive inflows materialize. The core argument across all of it, unless memory and semiconductors are no longer central to AI and there is no replacement in sight there is no reason to be worried about Korea, SK Hynix, or Samsung. Every dip is buyable and Milk Road Pro members are up massively on these trades, come join Milk Road Pro for just a dollar using the link below to get our full trades.

@BitcoinJesusETH @m0xt_ both

@MilkRoadAI @m0xt_ literally or figuratively?

FIGR is going to beat Q2 revenue estimates by 22%. And the EBITDA beat will be even bigger. Here is why: Revenue consensus is at $183M My estimate is $223M Figure pre-released Q2 marketplace volume: $4.26B Up 47% in a single quarter. That's $160M above the top end of their own guidance, and $300M above the midpoint Wall Street built its models on. Apply management's own guided net take rate (3.75%) to that volume, and you get $223M in revenue. At a 51% adjusted EBITDA margin, that's ~$114M in EBITDA vs. $92M consensus. A 22% revenue beat compounding into a 24% EBITDA beat. And here's the part most investors will read exactly backwards: The take rate is falling. The margins are rising. Take rate tells you what Figure earns per dollar of volume. It says nothing about what it costs to earn it. Contribution margin does. As Figure Connect (using external partners) grows from 56% to ~60% of volume, the take rate compresses. Maybe even below 3.75%. But Connect borrowers cost Figure almost nothing to acquire. The institutional partners do the marketing. So the mix shift looks like yield compression from the outside. From the inside, it's margin expansion in disguise. A business growing volume 47% QoQ. Not YoY, quarter over quarter. Running 50% EBITDA margins, guiding toward 60%. At $29.73, the stock trades at 7.4 times annualized revenue and 14.4 times annualized EBITDA on our Q2 estimates. Seven analysts cover it. The average price target is $52.86, 78% above where it trades today. The market is not pricing the earnings power this business is demonstrating right now. I've been adding FIGR to the Milk Road PRO portfolio ahead of the print. Full position and real-time moves inside ($1, link in bio).

@MelvinInvests Nice chart!

This is one of the most important charts for understanding where AI infrastructure costs are headed (Save this). Morgan Stanley just revised their cost per gigawatt estimates upward across every single rack architecture from Q2 to Q3 2026. For NVIDIA's most advanced systems, a Rubin Ultra cluster now costs $50 billion per gigawatt, up from $43 billion, a 16% jump in a single quarter. Two drivers are pushing costs higher. The first is memory and for Rubin specifically, HBM has gone from a low single digit percentage of rack bill of materials to approximately 25% roughly a tripling in share. HBM4 pricing on the Vera Rubin NVL72 rack has surged 435% compared to Blackwell, with memory alone costing over $2 million per $7.8 million rack. Rubin Ultra pushes this further, HBM4e swells to $1.53 million per rack on a $21 million total. The second is outside the rack infrastructure power, cooling, transformers, substations. Prices for next gen outside the rack costs at $16-19 million per megawatt, representing 41-49% of a GB300 cluster's total cost. Rubin Ultra Kyber pushes toward 600 kilowatts per rack and the electrical supply chain was simply not built for this. Now here are the stocks that benefit directly from this chart. Micron, SK Hynix Samsung are the only companies that can manufacture HBM4 and HBM4e at scale and with memory now tripling as a share of rack BOM, every new cluster deployment is a larger check written directly to them. On the power and cooling side, Vertiv, Eaton, and Schneider Electric are the primary beneficiaries of the outside the rack cost surge, liquid cooling systems, power distribution units and thermal management are now 41-49% of total cluster cost and growing faster than the compute itself. Broadcom and Marvell capture the networking layer, which remains 19-23% of total rack cost across most architectures. And NVIDIA itself sits at the center of all of it, Rubin Ultra at $50 billion per GW means every hyperscaler committing to a new cluster is committing to figures that would have seemed impossible two years ago. The AI infrastructure is getting dramatically more expensive to build at exactly the moment hyperscalers are racing to build more of it. Memory inflation and power bottlenecks cannot be solved by writing a bigger check, they require new HBM fab capacity and substation builds measured in years, not months. Make sure to follow me @MelvinInvests for more investment ideas.

@KyleReidhead I will take MSFT over apple

Everyone thinks Apple is losing the AI race because it skipped the AI capex But I think they are positioned perfectly to dominate AI (Save this) In fact, Apple is the best performer of the Mag 7 YTD, so the market is starting to figure it out I bought $AAPL earlier this year for the AI-demand device upgrade cycle, but I'm realizing they are positioned to dominate in something much bigger: Consumer Agents Apple just shipped App Intents, the framework that lets Siri take real actions inside any app, book something, buy something, complete a task, not just answer a question. My bet is that consumers will want to just talk to their agents without actually touching their phone "hey Siri" rather than picking up their phone, opening it, clicking on chatGPT app and then talking/typing Apple has a moat on devices with iPhone, Macbook and iPods. This is where consumer agents will be used. And because Apple doesn't have their own AI, they can become an aggregator of AI's, similar to Openrouter. Routing to the cheapest/most efficient models depending on the task plus, it can live locally on the device for privacy and speed. So Apple has spent no money on AI, yet is sitting in the perfect position to be the winner of how consumers use AI agents on daily basis. If Apple owns that layer, it gives them yet another (and likely one of their biggest) revenue streams. they become a platform that taxes the entire agentic economy, monetizing through tiered subscriptions and/or taking a cut of every transaction an agent completes inside the apps. And not only that, if the Agentic economy takes off through Apple, it will force the largest device refresh in Apple's history. Morgan Stanley says roughly 850 million iPhones can't run Apple Intelligence, and 1.3 billion can't run the new agentic Siri, out of about 1.4 billion active iPhones worldwide. Now of course, Siri is still dumb, so they have not achieved this yet. But the potential and roadmap is there. This is why I bought $APPL months ago and shared this with the members inside Milk Road PRO. I just shared a detailed update on my position inside the platform too. You can track my real-time portfolio and get all my live updates for just $1 (see link in bio)

We called Nebius, Credo, Bloom Energy, AAOI, and AMD before their big run ups. Don’t miss the next one, come join us for just a $1. milkroad.com/pro/?utm_mediu…

Jim Cramer strikes again, yesterday he called IBM a buy and this morning it crashed 25%, its worst single day decline since the dot com bubble (Save this). IBM pre announced preliminary Q2 results that came in well below internal expectations. Revenue was $17.2 billion versus $17.86 billion estimated up just 1% year over year. Software grew only 5%, consulting was flat, infrastructure was down 7% and the CEO himself said "this quarter we faltered." The company lost over $65 billion in market cap in a single session. The reason for the miss is what matters for the broader market and it is not an IBM specific story. CEO Arvind Krishna published a letter explaining exactly what happened. "In the last few weeks of June, IBM's enterprise clients abruptly shifted their quarterly capex budgets away from IBM software and mainframe contracts and toward servers, storage, and memory purchases specifically to lock in supply constrained infrastructure ahead of expected price increases. IBM said it "did not anticipate the magnitude of the capex reprioritization." This is the AI capex supercycle cannibalizing enterprise software spending in real time. Companies that were IBM's bread and butter customers for decades are now redirecting discretionary tech budgets toward GPU servers, HBM memory and storage rather than IBM licensing deals and mainframe upgrades. The broader implication is significant. Every enterprise software name, Salesforce, ServiceNow, SAP, Oracle's non-cloud divisions now faces the same risk that IBM just crystallized. When corporate IT budgets are finite and AI infrastructure is the new non negotiable priority, something has to give. Milk Road Pro is tracking every beneficiary of this AI infrastructure wave and every company that's about to get left behind. Come join for just $1 using the link below and position accordingly!

@MelvinInvests long nebius!

Nebius is putting every other neocloud to shame and this morning alone proves it (Save this). Nebius just signed a $1 billion compute deal with Reflection AI, running through 2029 and giving Reflection direct access to NVIDIA GB300 chips. This is the deal everyone should be paying attention to, because of who Reflection is. Reflection AI was founded by former Google DeepMind researchers behind Gemini's reward modeling and AlphaGo. They raised to a $20+ billion valuation in early 2026, backed by NVIDIA, Sequoia, and with direct US government support, their mission is to build America's open source frontier model, a Western answer to DeepSeek. In June 2026 they signed a $6.3 billion compute deal with SpaceX at $150 million per month for access to the Colossus 2 data center in Memphis. Now they are adding Nebius as a second major compute provider on top of that. The fact that Reflection chose Nebius, not AWS, Azure, Google Cloud for that relationship says everything about where the neocloud sits in the frontier AI infrastructure hierarchy. The CEO himself explained the ambition in a recent interview: "How many companies today provide hundreds of thousands of GPUs in a publicly available cloud? Well, there is the three hyperscalers and us." That is a direct challenge to Amazon, Google and Microsoft on their own turf, with the contracts to back it up. The India news this morning adds another dimension entirely. Nebius just posted two hardware infrastructure roles in Hyderabad, Field Technical Lead for Data Center Deployments and Technical Program Manager for New Data Center Launches. This is exactly how Nebius signals a new market entry, the Wales data center from yesterday was flagged the same way through a recruiter post before any official announcement. India would be Nebius's first footprint in Asia, entering the fastest growing AI market outside the US and China at a moment when data sovereignty regulations make domestic compute infrastructure a strategic priority for Indian enterprises and government contracts. This all comes on top of a buildout that is already staggering, a 310 MW AI factory in Finland, a 240 MW facility in France, a £1.7 billion UK expansion across three sites, Wales, a $20 billion contracted backlog from Meta and Microsoft, and $3.75 billion in convertible notes raised in March 2026. The only constraint on this company is how fast it can build. Bullish on Nebius and make sure to follow me @MelvinInvests for more underlooked AI oppurtunities.

Eli Lilly is going much HIGHER and it has almost nothing to do with weight loss drugs (Save this) Zepbound & Mounjaro are the reason the stock sits near ATHs but the AI platform $LLY is building under the drug business is what matters it started w/ Lilly Pod, a 1,000 GPU Nvidia Blackwell supercomputer at its Indianapolis HQ, the most powerful supercomputer owned by any pharmaceutical company on earth In January 2026 Lilly and Nvidia committed $1 billion over five years to a new AI co innovation lab running on Nvidia's BioNeMo platform and Vera Rubin architecture The lab pairs robotic wet labs with AI models in a closed loop: propose a molecule, synthesize it, test it, feed the result back in, and repeat at machine speed instead of human speed The next wave is data: Lilly CEO Dave Ricks has pointed out that Lilly alone holds data on roughly 3 million failed drug candidates, while the entire pharma industry has only ever produced about 4,000 approved drugs total No AI startup training on public internet data can replicate 150 years of proprietary clinical and safety data like that. This is their moat in an AI world And finally, wave 3 is about distribution: TuneLab launched in September 2025, giving outside biotechs access to AI models built on over $1 billion of Lilly's proprietary drug discovery data through federated learning Over 70 biotech partners are already on it That is Lilly starting to look like infrastructure the rest of the industry builds on, closer to what AWS became for cloud computing than a normal drug company Even better is that Eli Lilly is funding this bet entirely off cash flow from drugs that already work, Zepbound, Mounjaro, and the new oral pill Foundayo, so the AI platform is optionality stacked on top of a business that does not need it to succeed The drug franchise is already crushing, the AI platform is the option most of the market has not priced in yet and what will take $LLY much higher Follow me @kylereidhead for more analysis on AI, robotics and markets. Also, you can track my real-time portfolio (which includes $LLY) for just $1 at Milk Road PRO (see link in bio to join)

Infineon just made a quiet move on one of AI’s biggest bottlenecks: Power! It extends the $IFX thesis from the server rack to the entire data-center campus. (Save this) Infineon signed an MoU with LS Electric to co-develop three building blocks for next-generation DC data centers: Battery power-conversion systems, solid-state transformers, and solid-state circuit breakers. Infineon brings the power semis, controllers and software. LS integrates them into complete power systems. And LS Electric is a heavyweight, South Korea's leading data-center power supplier with roughly 60-70% of the domestic distribution market. In the clip below, Infineon President Adam White explains why solid-state transformers become critical as AI racks scale from roughly 120 kW today toward 500 kW and eventually 1 MW. At 1 MW per rack, data centers cannot simply push more power through today’s low-voltage systems. Doing so would require enormous currents, thicker copper cables, more cooling and create higher energy losses. Instead, power needs to be moved around the campus at a much higher voltage and only stepped down closer to the GPUs. That is where this MoU fits. The three technologies manage the power before it reaches the rack: → The solid-state transformer converts incoming grid power into the high-voltage DC network → The power-conversion system connects the data center’s batteries to that same network → The solid-state circuit breaker shuts down faults almost instantly to keep the system safe The takeaway is that this is a pattern that keeps repeating itself: Every few months, another piece of the AI power chain gets an Infineon partnership attached to it! You can track how I'm positioned across the 800VDC trade in Milk Road PRO. Link in bio.

We called Nebius, Credo, Bloom Energy, AAOI, and AMD before their big run ups. Don’t miss the next one, come join us for just a $1. milkroad.com/pro/?utm_mediu…

This Morgan Stanley chart is probably the most important single slide in investing right now (Save this). Morgan Stanley just raised its 2027 and 2028 hyperscaler capex estimates by 9% and 10% respectively and the new numbers: $1.23 trillion in 2027 and $1.4 trillion in 2028. Those figures include SpaceX terrestrial compute capex for the first time, a new entrant Morgan Stanley is now treating as a legitimate hyperscaler class spender alongside Amazon, Google, Microsoft and Meta. To put the trajectory in context: hyperscalers spent $261 billion in 2024 and now projected to spend $1.4 trillion in 2028 which is a 5x increase in four years. Now, who captures this spending and how could you benefit from all of this? NVIDIA is the most direct beneficiary, every dollar of AI compute capex disproportionately flows through GPU purchases first. Goldman Sachs estimates semiconductor sales revisions for 2026 are already up approximately 60% as a result of these capex uplifts and the $1.2-1.4 trillion spending level creates demand visibility essentially locked in 18-24 months forward. Micron and SK Hynix are the memory layer, every NVIDIA GPU requires HBM, high bandwidth memory stacked directly on the chip package and there is no substitute. Micron's HBM3E is now shipping at scale into the GB300 ecosystem and SK Hynix continues to dominate HBM supply overall. As capex scales toward $1.4 trillion, HBM demand scales in direct proportion and Micron is the most accessible US-listed pure-play on this dynamic. Broadcom captures the custom silicon and networking layer. Google TPUs, Amazon Trainium, Meta MTIA, every major hyperscaler custom AI chip program runs through Broadcom for networking silicon and packaging. As total capex grows and custom silicon share within it grows, Broadcom's addressable market expands on both axes simultaneously. Marvell captures the optical DSP and custom ASIC layer specifically the photonic engines and SerDes that connect GPU clusters together at scale. Fiscal 2027 revenue is tracking toward $11 billion, $15 billion guided for fiscal 2028, driven almost entirely by hyperscaler AI commitments already in backlog. Vertiv and Eaton capture power and cooling infrastructure, every GPU cluster requires thermal management and power distribution at roughly $0.30-0.40 of infrastructure spend per dollar of compute. At $1.4 trillion total capex, that translates to $400-550 billion flowing into power and cooling infrastructure over the next two years. The infrastructure suppliers get paid regardless of whether AI ROI materializes while the hyperscalers bear the risk. That asymmetry is the core reason the picks and shovels trade remains compelling even at current valuations. Milk Pro Subscribers are up massively on these trades and come join us using the link below to get our full AI trades and today is the last day to get 33% off.

We called Nebius, Credo, Bloom Energy, AAOI, and AMD before their big run ups. Don’t miss the next one, come join us for just a $1. milkroad.com/pro/?utm_mediu…

We called Nebius, Credo, Bloom Energy, AAOI, and AMD before their big run ups. Don’t miss the next one, come join us for just a $1. milkroad.com/pro/?utm_mediu…

@MilkRoadMacro @m0xt_ $MU is the better memory play

SK HYNIX HAS CRASHED 20% IN JUST TWO DAYS. But here's why we are not worried. 👇 Our analyst holds SK Hynix in his portfolio so here's his latest update: @m0xt_ added a speculative position last week betting on SK Hynix's Nasdaq debut. Last night, SK Hynix was down 15%. Here is what he's doing about it: SK Hynix listed on Nasdaq at $149 per ADR last Friday and price jumped 16% that day. 7x oversubscribed. One of the biggest US listings in years. He wasn't buying Hynix directly. He was buying SK Square which owns roughly 20.5% of Hynix and trades at a deep discount to what that stake is worth. The thesis was a dual discount closure. First, a Korea discount on Hynix itself as dollar-denominated Western capital finally gets access. Second, the holdco discount on SK Square narrowing as the underlying value becomes undeniable. The US listing was the catalyst that could trigger both. He had already held SK Square as a long-term position. But ahead of the Nasdaq debut, he added more. That add was speculative: He was betting on a short-term re-rating as the listing closed the discount fast. He wanted exposure to that move specifically. Then over the weekend, US-Iran tensions escalated. Korean retail investors were already sitting at extreme leverage levels. When macro fear hit markets, there was forced selling. Hynix fell 15% in Seoul. SK Square fell 17%. The short-term trade he added for did not play out. Now he has two options. He can cut the speculative portion, take the loss and free up the capital. Or he can convert that tranche into the long-term thesis and hold alongside my original position. He's doing the latter. The long-term Hynix thesis did not break over the weekend. The holdco discount, the Western capital access story: none of that was touched by Iran. What changed was Korean retail sentiment and leverage, a flush I could not have predicted. The setup he entered on is still there. The real cost is that his capital is locked longer than planned and his available cash is already low. That limits my ability to act elsewhere. That is the honest tradeoff I am accepting. What I am watching: whether SK Hynix continues to compound as a business. And at current prices, he thinks the probability is quite high. Hynix is expected to generate roughly 150 billion in profits this year, priced at around 7x earnings. Memory demand is structurally tied to AI infrastructure spending and supply shortages are likely to persist through 2030 as demand is already outpacing supply before robotics even becomes a meaningful driver. That wave is starting. If Hynix executes into that backdrop, the share price follows on both exchanges and SK Square's discount closes on its own. That is the long-term thesis. The premium spread was the short-term trade. He is no longer in the short-term trade. These are the types of updates you'll receive if you're inside Milk Road PRO. Our analysts give daily updates on most of their portfolio holdings. Even if the trade doesn't work, they're the first ones to take accountability just like @m0xt_ has done above. You can check all our winners and losers inside Milk Road PRO. For the next 12 hours, there's a 33% discount on PRO. Grab it before the offer expires (link in first comment below).

We called Nebius, Credo, Bloom Energy, AAOI, and AMD before their big run-ups. Milk Road is running a HUGE 33% promo right now. Don’t miss the next one, come join us: milkroad.com/pro/discount/?…

Nebius opening a data center in Newport, Wales is a small headline but the company's trajectory is not (Save this). Nebius is a neocloud and it builds and operates its own NVIDIA GPU data centers and rents that compute capacity to AI labs and enterprises at scale. The company is actually the renamed remnant of Yandex, the Dutch holding company of Russia's search giant, which sold its Russian operations for $5.4 billion in 2024, relisted on Nasdaq as Nebius Group under the ticker NBIS and pointed its entire engineering team at the AI compute buildout. In the 18 months since that relisting, AI cloud revenue grew 841% year over year in Q1 2026 alone. NVIDIA invested $2 billion directly into the company, Meta signed a contract worth up to $27 billion over five years for dedicated compute capacity. The company went from essentially zero revenue to a $20 billion contractual backlog in 18 months. The European buildout is what makes the Wales site meaningful in context. Nebius is executing one of the most aggressive data center construction programs in Europe, a 310 MW AI factory in Finland worth over $10 billion, a 240 MW facility near Lille in France and a £1.7 billion UK expansion across three NVIDIA Blackwell Ultra sites by 2027. Newport is the newest addition to that UK footprint. The strategic logic is data sovereignty, every European government is prioritizing the ability to run AI workloads on infrastructure domiciled in Europe, subject to European law, not American jurisdiction. Nebius, headquartered in the Netherlands, is positioned as the dominant European neocloud at exactly the moment that positioning has maximum political and commercial value. The company is targeting over 3 gigawatts of contracted power by end of 2026 dedicated entirely to AI compute. The only real constraint on growth is how fast they can build. Bullish on Nebius and our subscribers are up massively on this trade, come join us using the link below to get our full trades and we are having a 33% of just for today.