Matter

$AMD Lisa Su talks about how Anthropic is likely a multi-generational customer of AMD GPUs beyond MI455X. It’s too much work to get teams integrated and just to adopt for a single generation. Implies the deal is beyond just initial 2 GW.

On the flip side, if depreciation schedules remain unchanged, the residual value of GPUs go higher and generate revenues at much higher incremental margins after fully depreciated

Incredible that even ~3-4 year old GPU models ($AMD MI300, $NVDA H100) rental prices are still going up ~30% per $DOCN. Possibly +ve to useful life estimates for GPUs likely beyond ~5 years, reducing D&A expense + boosting margins for hyperscalers. Disproves AI overbuild fears

@SemiAnalysis_ Also Anthropic’s own compute constraints are clear with recent $SPCX and now $META providing monthly compute rentals via excess capacity. So using $AMD MI455X would only make sense for Anthropic. Has anyone heard of SpaceXAI being a potential customer of Helios also?

@SemiAnalysis_ This doesn’t mean AMD has won Anthropic yet, but it’s def one of its highest priority accounts for software engineering resources, increasingly the likelihood of a win. Anthropic also has a history of focusing on heterogeneous XPUs (I.e. models running on Trainium, TPU)

ANTHROPIC WILL BE AN AMD CUSTOMER, ACCORDING TO THE PUBLIC GITHUB OF AMD’S SENIOR DIRECTOR OF AI 🚨🚨 We explain the GitHub code and nuances below👇️ 1/4🧵

@jukan05 The use of Nangate 45nm matters enormously. It is an open academic standard-cell library, not a production 2nm or 3nm foundry environment. A 45nm design faces dramatically less complexity

EDA is cooked.

7/7 Bottom line: LTAs protect the downside. HBM/AI DRAM drives the upside. NAND moving from cold storage toward warmer AI memory/storage tiers adds optionality. The next memory downturn can still happen, but it may not look like the old cycles. $MU $SNDK

1/7 Memory LTAs may be more important than the market appreciates. The key change: these are not old buyer-friendly LTAs. They increasingly include volume/price commitments, cash deposits or LOCs, and multi-year terms. That could structurally dampen memory cyclicality.

6/7 HBM makes this more interesting. Micron is not just a commodity DRAM/NAND story anymore. HBM4/4E/5 require tighter customer qualification, advanced packaging, custom base dies, power/thermal co-design, and scarce qualified supply. That should support +ve pricing durability

5/7 The biggest Micron implication is not that downside disappears. It is that trough gross margin and EPS may be structurally higher than prior memory cycles. If true, the old “peak earnings deserve a very low multiple” framework may be too punitive.

4/7 Also protection improves over time. As customers purchase under the contract, remaining obligations decline. The same deposit/guarantee then covers a larger % of the remaining contract value. So LTAs may be most powerful later in the cycle, when downturn risk is higher.

3/7 I would give much more credit to cash deposits/LOCs than potential legal damages. Legal recoveries may be possible, but they are slow, uncertain, and relationship-sensitive. The immediate financial commitment is the real first-loss protection investors should underwrite.

2/7 The breach math is different now. A customer cannot simply compare contract ASP vs spot ASP for DRAM/NAND They must compare contract ASP vs spot ASP + forfeited deposit/LOC + legal/friction costs + future supply access risk. That raises the hurdle to walking away.

@jukan05 If this is true then doesn’t GUC (Global Unichip Corp) benefit as they do backend design, and are strategic partners of TSM who is the top shareholder? They already do the backend design for Microsoft MAIA XPU

Rumor: Google to shift to a COT model starting with TPU v9, downgrading the role of MediaTek/Broadcom. This would be a major negative for Broadcom and MTK. According to this rumor, starting with TPU v9, Google plans to move to a model similar to AWS’s Annapurna, where Google directly places wafer orders with TSMC. In other words, Google would be doing COT directly. What does that mean? Google would become TSMC’s direct customer, while Broadcom and MediaTek would be downgraded to IP support / design service roles. Ownership of the chip and control over supply would shift toward Google. In other words, Broadcom/MediaTek’s per-chip margin would fall, if this is true. Instead of recognizing the full chip ASP as revenue, the model would shift toward design service fees / margin. That makes me more worried about Broadcom than MediaTek. $AVGO

@jukan05 If Intel is targeting 30% price hike this yr, can we assume AMD does on similar level, plus strong unit growth, and add share gains on top due to Intel near capacity constraints = AMD Datacenter CPU rev growth could be 50%+ this yr? (Vs street ~40%)

A very spicy piece of information here. A must-read for everyone (especially if you're interested in the CPU shortage). - Intel has carried out two rounds of CPU price hikes since the start of 2026. The first round in February was 10–15%, and the second on March 16th was 15%, bringing the cumulative increase to ~20% vs. January. - The full-year target is a cumulative 30% price increase, aimed at recouping investment and meeting capital market expectations. - A third round of price adjustments is reportedly planned for May. - Server CPU demand is strong globally, with orders already up 50% YoY vs. 2025. - Chinese CSPs in particular are placing additional orders — Tencent is scaling its server CPU procurement from ~300K units in 2025 to a 2026 target of 900K–1M units, deploying them toward AI infrastructure buildout. - Alibaba has the largest procurement scale among Chinese CSPs, and is also expected to grow 30–35% YoY. - Consumer CPU shipments are also picking up, driven by local AI PC demand, with a projected ~20% YoY increase in shipments (though I find this one a bit dubious). Intel's consumer CPU share remains stable at 65%. - Intel's production line utilization has reached 95%, but is not yet physically at full capacity. Server market deliveries are smooth (China Q1 orders +60%, deliveries +27%), but on the consumer side, despite orders doubling (+100%), deliveries actually declined 3.8%. This is due to PTL yield issues and new product supply being prioritized for North America — Intel is managing this by extending lead times to 3 months. - Intel's consumer CPU market share remains stable at 65%. - On the server side, Intel's share has dropped from 65% in 2025 to 50%, primarily due to AMD's stable supply and superior price-performance. - A particularly noteworthy point: the CPU-to-GPU ratio in AI servers has risen from 1:12 to 1:8, with a target of 1:4, meaning the share of CPU procurement is increasing. Very important. - Cloud providers' AI servers generally have excess compute capacity, and there are no plans for large-scale traditional server expansion for the time being. - AI applications are expected to drive ~10% of incremental demand for consumer CPUs, with long-term growth dependent on product usability. $INTC $AMD $ARM

@shanaka86 It makes sense inventories are ramping also as they’re ramping Blackwell Ultras (B300s) in 2HCY25. Inventories always tend to ramp early in a product launch. These inventories will then be recognized once OEMs can ramp up rack capacity

@shanaka86 This doesn’t spell “doom”. It’s known there’s an upstream/downstream mismatch between the OEMs who build racks, and NVDA’s supply of chips. These OEMs/ODMs are scaling up their capacity however. NVDA and others therefore hold finished goods inventories

The $33.4 Billion Shell Game: How Silicon Valley’s Greatest Fraud Unravels in 40 Days Nvidia reported something devastating yesterday that nobody caught. Days Sales Outstanding jumped to 53 days. Historical average: 46 days. That seven day difference represents $10.4 billion in revenue Nvidia collected on paper but never in cash. Here’s what’s actually happening: Nvidia invests in AI startups. Those startups buy cloud services. Cloud providers use that money to buy Nvidia chips. Nvidia books it as revenue. But it’s the same money going in circles. The proof is mathematical: • Accounts receivable: $33.4 billion (doubled since last year) • Inventory: $19.8 billion (rising during a “shortage”) • Cloud commitments: $26 billion (doubled in 90 days) Total capital trapped: $79.2 billion. Total cash generated last year: $64.8 billion. They’ve trapped more money than they’ve ever made. The smoking gun: Inventory rising 32% while claiming “insane demand” is impossible unless those chips aren’t actually selling. You cannot have shortage and surplus simultaneously. Basic physics. Operating cash flow is only 75% of net income. Healthy companies generate 100% or more. That 25% gap? Fake revenue that will never become real money. What happens next: December 2025: Aging schedule reveals truth February 2026: Last chance to exit April 2026: First receivables writeoff October 2026: Full unraveling begins Stock price today: $140 Price after writeoffs: $70 This isn’t speculation. It’s accounting arithmetic. When receivables age beyond 60 days, writeoffs are mandatory under GAAP. Nortel did this in 2001. Lucent in 2000. Both went to zero. The mechanism is identical: circular vendor financing disguised as growth. Verify yourself: Check any tech company’s DSO. Above 50 spells doom. The countdown has begun.​​​​​​​​​​​​​​​​ Read the super deep dive article here - open.substack.com/pub/shanakaans…

@smallmidcaps @polemicuss @moazzam0_reddit @AndrewP64090323 I think NRP can only be used more locally when the distance is minimal. If you put a NRP liver on SherpaPak it will still go thru ischemic damage (albeit to lesser extent than a traditional ice box)

@MatterGlobal @polemicuss @moazzam0_reddit @AndrewP64090323 Other machine perfusion companies aren't a threat in the short-term. The 3rd party NRP companies using the chaper portable devices like sherpapak and arranging logistics is a more immediate threat. Why else is NRP growing?

Last post I ever make about $TMDX Below are the full numbers. TMDX is expected to grow liver revenue 96% (vs. mkt at 79%), heart 72% (vs. mkt at 30%) and Lung 58% (vs. mkt at 39%). They have now 85% mkt share in DCD liver. They are quite literally the DCD liver market

@polemicuss @moazzam0_reddit @smallmidcaps @AndrewP64090323 Agreed. The moat is indeed the end to end service. One liver tx expert call stated that their hospital used 70/30 OCS vs OrganOx for their machine perfusion (90-95% of liver cases). Reason being OCS has NOP (full service) and OrganOx is not portable via plane

@moazzam0_reddit @smallmidcaps @AndrewP64090323 @MatterGlobal I think the moat of TMDX is not as much the warm perfusion tech as the service they've built around it. Hospitals are not in the business of logistics.

@polemicuss At 10x+ was pricing towards near perfect execution for sure (I.e steady QoQ growth). But even at those levels it still wasn’t pricing in 10K Tx by 2028 and ~30% Ebitda margins at scale. Near term execution will be key for mkts to gain confidence in their mid term targets again

@MatterGlobal I agree - all i wrote is in the context of valuation. TMDX at 10x+ was insanity