The Silicon Stack Podcast | Ep. 2 Featuring @PaulBPrager I sat down with $WULF CEO Paul Prager. The Naval Academy grad who has transformed TeraWulf into one of the fastest-growing pure-play AI/HPC data center platforms in the U.S. We discuss: • The 20-year Anthropic deal • Statewide data center moratoriums • Power generation (BTM & FTM) • Overseas expansion plans

ANTHROPIC: OUR GROWTH STRATEGY “Compute. We intend to secure further compute beyond our existing commitments, expand our own data center footprint to secure additional long-term compute, improve compute efficiency across our multi-chip compute architecture, and produce further algorithmic advances in model efficiency”

ANTHROPIC S-1: “OUR APPROACH TO COMPUTE” $NVDA $GOOGL $AMZN $AMD

Morgan Stanley: Buy Powered Shell Providers We know analyst Stephen Byrd is bullish, but is he too bullish? $WULF $CIFR $MARA $HUT $GLXY $APLD $RIOT

BofA sees $NBIS exiting the quarter with 675 MW of connected power and ~525MW of active power, representing a meaningful step-up from the 170MW active at the start of 2026. The primary focus will be Nebius’ ability to achieve its year-end target of 800 MW to 1 GW of connected power, particularly potential delays at Vineland.

$NVDA CEO Jensen Huang on neocloud concerns: “This is gonna get solved because demand is so good.” He sees the real issue as multi-month lag from GPU delivery to revenue and the massive capital float required. Huang wants to shorten the supply chain so operators can stand up data centers faster.

Data center sentiment is as bad as it’s ever been $IREN $CRWV $CIFR $NBIS $WULF $GLXY

JP. Morgan: The global Server market is estimated to expand to $506 bn in 2026 from $284 bn in 2025, implying +78% YoY gowth. Estimated 31% CAGR from 2026-2030 $NVDA $DELL $SMCI $HPE $ANET $CSCO

Morgan Stanley on $WULF - “Addressing the biggest misconceptions” PSPs are often confused with neoclouds. 1. PSPs enter into 15-20 year leases, rather than 2-3 years for neoclouds 2. Powered shells built by PSPs are less capital-intensive ($9-14/watt, rather than $30-50/watt for the neoclouds) 3. PSPs take no chip risk, they are simply landlords We believe it is instructive to compare PSP stocks with real estate stocks. Many real estate investors use discount rates and cap rates that are much lower than those embedded in PSP valuations. For example, in our base case for WULF (441% upside to our price target), we use a 9% equity discount rate and 7% terminal cap rate on contracts that we expect the company to sign over the next few years.

SPOT COMPUTE TRADING AT 3-4X LONG-TERM PRICING: Spot and short-term deals are trading at $40-50M/MW: ($SPCX $NBIS $CRWV) Long-term IG deals sit between ~$10-16M/MW: ($IREN $ORCL $NBIS) Near term scarcity is extreme

JP. Morgan on $SPCX neocloud business: “CEO Elon Musk addressed neocloud pricing dynamics with the belief that value of compute per watt will increase in tandem with the increase in intelligence per watt. As more compute capacity comes online and models improve, intelligence per watt will improve, ultimately leading to higher pricing on neocloud contracts.” Current short-term neocloud contract pricing north of $50/W; pricing to increase alongside intelligence per watt; positive pricing datapoint for neoclouds but capacity build could displace other providers.

Senior Vice President at Cast believes the long-term risk for neoclouds like $CRWV is that hyperscalers are aggressively developing custom silicon (internal accelerators) and self-building at massive scale, which may eventually reduce their dependence on third-party GPU lessors after 2028. “It's early on in the custom silicon game relative to the valuation that we're giving GPU providers in neoclouds today. I think if you're assessing how the AI inference could erode the market share, it's still the biggest long-term threat, but it shrinks CoreWeave's share of AI compute rather than the GPU market's absolute size, at least for the rest of this decade. The threat is concentrated in exactly the place that CoreWeave relies on, hyperscaler overflow demand and inference, which is both the fastest-growing workload and where older GPUs go for their second life. Frontier training for labs that don't own chips is probably the most protected part of that market.” The expert from this post is a Vice President focused on AI ecosystem and cloud alliances at an AI infrastructure company, he has spent the past decade working in cloud and AI infrastructure, including enterprise sales roles at Amazon and Microsoft.

AWS currently prices B200 capacity blocks at roughly $12.36 per GPU-hour and B300s at $14.04, implying a neocloud discount of roughly 30%-45% on comparable current-generation hardware $SPCX $CRWV $NBIS $IREN Source: William Blair (9/30/2026)

Expert call: $NBIS Former global strategic partnership lead at Nebius says Nscale’s AI factory is closer to $IREN than Nebius or $CRWV: “Nscale at the moment is investing considerable amount of resources in the AI factory scale, especially if we are talking about the deployments in the U.S. as well as in Norway. By saying that, how they position themselves in comparison in the space. At the moment, they are not at the same level of CoreWeave or Nebius, they have still quite substantially time to go. They position themselves closely to IREN at the moment. Even though IREN is more specialized than, for example, Nscale with acquisition of Mirantis, this is a more complete offering for enterprise solutions, in this case. Nscale, they are moving towards their direction. They become a fully integrated neocloud with specialized GPU as a service and probably moving forward toward inference as a service. At the moment, though, like I mentioned before, because of limited capacity, so we are still small in space, even though we have significant contracts also with Anthropic. It takes time to deploy, at least normally standardized time for deployments is going to be 18 months. It means that they are spending entire 2026 in actually preparing the facilities for the deployments, and then they are going to be start in the second half of 2027 to start to have more commercial deployments, more bare metal that's going online. A key for Nscale’s strategy was not only to secure the contracts for the capacity. We are talking about almost 500,000 GPUs, but also because of the IPOs.”

BofA: European power demand could more than double by 2030E 🇪🇺 We forecast European data center power demand to rise from 17.6 GW in 2026E to 43 GW by 2030E. By 2035E, power demand could reach 111 GW. $NVDA $GOOGL $NBIS $CRWV $BE $META

$NBIS: “Execs say all 2027 capacity is sellable now” with BofA seeing 3-6 month deals at $44/watt annually. $40-50M/MW is the standard for 2027+

Goldman Sachs: The token demand curve for Robotics will likely create another lever to long-term data center demand dynamics $CRWV $SPCX $NBIS $GOOGL $NVDA $IREN

Greenwich just published a 15-page deep dive on the $IREN x Microsoft contract: Key takeaways for IREN: Construction & Delivery / Acceptance & Performance Risk • Horizons 3 and 4 cannot be delivered until Horizons 1 and 2 are operating as a single cluster. • Microsoft can terminate a tranche (and recover the upfront) if the minimum GPU quantity is not delivered for acceptance testing within the delay window. • A documented NVIDIA failure to deliver GPUs is an excluded delay (extends the window), but other delays can trigger the late-delivery exit. • Before acceptance, Microsoft can reject a tranche only if it fails the agreed acceptance criteria (and only within 5 days). • After acceptance, IREN’s fees are “non-refundable, non-cancellable” through 2031. • Service-level shortfalls and late deliveries after acceptance mainly earn delay credits or sit in a redacted service-level exhibit, they do not generally allow Microsoft to exit the whole contract. • The uncredited upfront is refundable only in narrow cases (Microsoft terminates for uncured material breach other than late delivery or service-level failure, or IREN ends the agreement under section 7.g).

Truist: We view the softness in $CRWV and $NBIS shares as a buying opportunity “Amid the weakening sentiment we believe investors should remain focused on the underlying fundamentals of these two AI cloud providers, which are only strengthening.” • Compute supply constrained through 2028 • Pricing power continues to expand • $40-50M/MW on shorter-duration deals • $10M/MW on long-term deals

Ken Griffin’s Citadel disclosed a 5.5% stake in $CIFR 😅 The 13G is mostly Citadel Securities market-making inventory, not a directional bet by the fund.