$NBIS vs $IREN: people are still beefing, and considering how the executives of both companies interact, it's going from a waste of time to actually stupid. Just today, under Daniel Roberts' LinkedIn post on how IREN funds its GPUs, Nebius' VP of Sales commented "don't give away all the secrets" (with two rockets lol). When I interviewed @romanchernin, it dawned on me that the companies themselves can't understand why their retail investors are fighting, when they would rather build relationships. $IREN investors always pointed out that Nebius doesn't have the power. $NBIS investors always said "but we can get it, and the team matters more." That difference in assets and skills is exactly what makes a collaboration so win win. $IREN has land and grid access. $NBIS, I believe, has the skills to build faster and increase the $ / MW that can be made out of it with a world-class cloud offering. 1 + 1 = 3. It's not a zero sum game. When you start a company, you do NOT want two CTOs or two CEOs as co-founders. You want aligned values and goals but DIFFERENT assets. Partnership or even M&A, I don't think anything is off the table. My view that $NBIS is the better investment is unchanged, but the more time passes, the more convinced I am that these two will end up working together. Maybe they already are behind the scenes. The setup makes too much sense and now the behaviour of the people involved also suggests interest on a human level.

SemiAnalysis: Datacenters SemiAnalysis sees 51.2GW of datacenter capacity being added in 2027. "Year-to-date, slightly under USD 1T of binding / non-cancellable datacenter leases have been signed – that’s key to supporting our aggressive 2027-28 forecast." SemiAnalysis claims Brad Gerstner's @altcap claim of only 25GW being built in 2027 is incorrect, based on satellite images showing >24GW of datacenters in advanced stage of construction, all to be delivered in 2027.

Yep, I can confirm this. A friend engineer at Reflection AI told me that even though they signed a compute deal with NBIS, they don’t know when they’ll actually get it. Signing a deal is easy; bringing compute online is an entirely different story.

$IREN: H2 Root Cause Analysis Update I got more information tonight confirming that the issue stems from either MSFT AI model research team not communicating workload requirements to MSFT HPC systems team or that MSFT HPC team incorrectly translated those workload requirements to networking requirements. MSFT HPC systems team communicates the requirements to IREN and defines the acceptance tests. The miss was both in specs provided to IREN and acceptance test for H1. This is not an error discovered while bringing up the network core to combine H1 and H2. Yes if IREN had a principal networking engineer with deep knowledge of AI workload requirements, they could have prevented this but it's MSFT's responsibility to give the the full requirements as they own the model and platform layers. Expectation is that H1 continues billing and new specs are implemented as maintenance over a 4 months. For H2, this is a design change that IREN has to respond to which will incur a 45 day delay. H3 is not most cabled at all so will not be delayed as much and be delivered at the same time as H4. H4, not cabled at all, will incur no delay and be delivered in Dec. H1-4 will all be billing exiting 2026. For those wondering why cabling can result in a 45 delay and worrying that a 45 day must mean structural changes in liquid cooling or power distribution, you can watch this youtube video to understand how time consuming cabling can be. Credits to @pietrovl on good video pointing out how complex cabling can be. Most likely, IREN is still using MPO since EBO came out just this July and hard to supply chain.

Coreweave’s CRISIS is a GPU pricing problem. $CRWV fell 23.7% this week asthe cost of insuring its debt hit a record (cc @amitisinvesting), implying roughly a 60% chance of default within five years. But looking at the underlying assets and contracts, it's clear that the business of GPU financing is fundamentally uncertain. (1/)

微軟退出英國 neocloud Nscale 在 Essex 的資料中心案，表面上是一則合約新聞，其實是 AI 工廠最現實的一道關卡：GPU 買得到，機櫃做得出來，但電接不上。 照 Telegraph 和 Bloomberg 的報導，主管機關告訴 Nscale，Loughton 這塊地要到 2030 年代才有電可接，原本預計 2027 年就要上線。微軟是這個場址唯一公開的客戶，原本要在這裡租約 2.3 萬顆 NVIDIA GPU。電一延就是好幾年，這個案子等於判了死刑。但要講清楚，這不是微軟跟 Nscale 分手。英國案只佔微軟合約的一小部分，葡萄牙 Sines 和德州的合約都還在，微軟 4 月才在葡萄牙加碼 6.6 萬顆 Vera Rubin。被丟掉的是那塊沒電的地，不是這家公司。 真正讓人不安的是 Nscale 本身的結構。它正準備在紐約上市，代號 NSCL，Telegraph 說估值最高上看 350 億美元。S-1 寫合約總值約 1,034 億美元，其中微軟約 438 億、Anthropic 約 446 億，兩家合計約 85%。但到 8 月底，真正在跑的只有約 26 億。剩下那 1,000 億，要等電接上、機櫃到位、準時交付才會變成營收，挪威旗艦機房原定 30MW，目前只上線 6.4MW，進度已經落後半年。更關鍵的是，S-1 寫明 Anthropic 那批要是交付延誤，對方可以不負責任地終止。換句話說，大型雲端業者把產能外包給 neocloud，延期的風險就一起推了出去。微軟自己蓋機房，延期得自己吞；租 neocloud 的產能，延期就可以退。風險全落在 neocloud 身上，而它們的錢大多是借來的。 這也是為什麼上週整個 neocloud 族群先跌了一輪。CoreWeave 單週跌近 25%，債券殖利率破 16%，5 年期 CDS 破 1,000bp，Nebius、IREN、Applied Digital 也各跌約 9 到 10%。Nscale 上市被當成市場還肯不肯借錢給 AI 基建的試金石。這時候再看 NVIDIA 跟 Nscale 的關係就更微妙了：NVIDIA 同時是它的供應商，也是客戶（12 億美元保留算力）、投資人（10 億美元，預計 11/16 交割）和租約保證人（約 8.6 億美元）。這是典型的循環融資，Nscale 上市要是卡住，NVIDIA 這條線的曝險就會被放大檢視。Dell 也在線上，Nscale 的 GB300 NVL72 機櫃是 Dell 整櫃交付，場址延期，機櫃就得延後出貨或改送別處。 當然也可以問另一個問題：微軟真的只是因為沒電嗎？Reuters 提到，大型雲端業者的第三季財報暗示資料中心案延期、資本支出成長放緩。缺電是真的，順勢砍掉最弱的一塊產能也說得通，兩件事可以同時成立。但我不認為這代表 AI 需求轉弱。Anthropic 8 月才簽下 446 億美元的合約，需求方還在搶算力，卡住的是供給端的電和資金成本，不是 GPU 沒人要。 所以這件事最大的訊號是：沒有自備電源的場址，在 2026 年已經很難賣了。Nscale 在西維吉尼亞給 Anthropic 的園區走的就是自己發電，最高 8GW，2028 年上半年先做到 2GW。受益的是燃氣渦輪、燃料電池、變壓器和開關設備這條線，美股的 GEV、VRT、ETN，台股的台達電、華城、士電、中興電、亞力，吃的都是電網排隊這個結構性題目。但它們跟 Nscale 沒有直接合約，主要客戶在台灣和美國，不要把英國一個案子直接對到它們的營收。鴻海、緯創、廣達這些機櫃代工廠也一樣，就算有經手 Dell 這批，頂多是出貨延後或改送別的場址，訂單不會消失。 看 neocloud，別只看合約總值寫多大。要看每個場址什麼時候真的通電，有多少合約真的在跑。

Between 2020 and 2022, approximately 960 SPACs went public. Of these, about 365 SPACs (38%) failed to find a merger partner and liquidated. Additionally, 277 SPACs (32%) liquidated by mid-202 Of the remaining 318, only 39 are still “alive” Of the 39 only 11 are over their SPAC valuation.

This is actually funny as fuck There's multiple sites with item that have had issues, not 1 TogetherAI was downgraded in ClusterMAX because of reliability issues. Vipul is not even guju and I know him and like him but our feedback and downgrade is because of quality issues. We remain fair and objective. Nebius downtime and reliability issues were much less. There's not a single cloud that doesn't have downtime, the question is how much, how often, etc. Your post has so much misinformation even in the first section no point in addressing the rest

China does not fuck around: "Xi Jinping wants those who benefited the most from China’s economic ascent to increase their contributions back to the state, embarking on what could be the most ambitious global tax hunt in history."

MAI-Code-1.1-Flash is going local with on-device model calls inside Github Copilot. High-quality agentic coding on your device, with no inference charge for local model calls.

$IREN: Addressing the Semi-Analysis FUD on Reliability What's the Issue? I heard of these issue in August and the only customers that $IREN has had long enough to report "multi-day power outages, network upgrades, storage failures, air quality controls leading to link flaps, XIDs everywhere" are FireworksAI and TogetherAI at Prince George. Mackenzie is undergoing GPU commissioning as we speak (1) so I don't know why @SemiAnalysis_ mentions Mackenzie having customer problems already when there are no customers there yet. What @dylan522p says is true but convinently omitting critical industry context makes this a cleverly disguised Semi-Hit-Piece. I had heard that TogetherAI filed this complaint to IREN and Dylan was able to obtain the information from TogetherAI's CEO. I was also told that Dylan is personal friends with TogetherAI's CEO and are from the same region in India and speak the same Indian dialect which is how Dylan is able to get the info. I will show that these are industry wide problems and the affected customers renewed despite these issues. First Industry Fact First, the most important fact that @SemiAnalysis_ conveniently leaves out is that on August 27, 2026, it was confirmed that “Together AI and Fireworks AI have both renewed and expanded” (2). Despite all these problems, why did TogetherAI and FireworksAI renew? Let's look at the rest of the industry facts. Second Industry Fact Second, I will address the most alarming sounding issue: multi-day power outages. Down-time happens to everyone in the industry. @SemiAnalysis_ Platinum Cloud, $NBIS, had a outrage spanning August 19-August 20, 2026 (3). It's not narrated by SemiAnalysis in hit piece voice but the results are similar: "a storm-related event affected the building of the data center facility hosting the us-central1 region, taking the facility’s building management system offline and shutting down the chilled-water cooling loop. With active cooling lost, the data halls overheated within roughly two hours, and servers, network switches and rack power systems began shutting down on thermal protection. This caused a region-wide disruption of us-central1 services: GPU and CPU compute, external network connectivity for virtual machines, object and block storage, Managed Kubernetes, Token Factory, managed platform services, and the regional API and console endpoints" (3). "Storm-related event" makes NBIS outrage explained. SemiAnalysis's biggest bias is that they conveniently leave out the reason for IREN's outraged: BC Hydro's grid had a brown out. That's why TogetherAI renewed, not because they were okay with unexplained, random multi-day power outrages. At the time, IREN did not have enough backup generators in BC since generators produce significantly more pollution than a hydro powered grid and are hard to obtain licenses for. By March, IREN was able to obtain license to put generators for all of PG and multi-day power outrages will no longer be a problem. IREN has designed Childress and SW1 with full N+1 backup power since MSFT which requires a higher uptime which is why SemiAnalysis says those sites "look much better". Third Industry Fact Third, let's address the rest accusations: "network upgrades, storage failures, air quality controls leading to link flaps, XIDs everywhere". Using fancy technical terms without explaining them is a hallmark of a hit piece, so I'll give definitions. Link flaps means a network connection is going in and out. XID is an error code logged and reported by Nvidia GPUs. Link flaps are common: SemiAnalysis Platinum Cloud $CRWV (4), Silver Cloud Lambda (5), $MSFT (6) all had 2hr15m (4), 5hr14m (5), 4h57m (6) network out rages due to link flaps this year. Why? Because when optical networking links and transceivers collect dust particles the connection can be interrupted. SemiAnalysis conveniently uses "air quality controls" to make it sound even more scary. "XIDs everywhere" sounds scary too but they have a much larger frequency on GPUs than other compute types. $META had a paper showing that they had 419 XIDs in 54 days on their infrastructure. No other AI Cloud is this transparent about reporting problems but $META blog is one of the most honest tech blogs out there bcause META culture is "Move Fast and Break Things". Network upgrades or other control/management plane upgrades do cause problem from time to time. AWS had a DynamoDB DNS defect (8) and GCP had a policy update trigger crashes globally (9). Storage outrage is the most outrageous accusation as those are the most common across both GPU and CPU Clouds. In a completely separate incident from the storm-incident, NBIS had a storage outrage in 2025: "we had an outage of the S3 service in the eu-north1 region triggered by a steady increase of migration traffic accompanied by unexpected behavior of the program" (10). Storage downtime are not typically reported these days but in the past generation, they use to more commonly reported by GCP (11), Azure (12), AWS (13). Storage is check pointed and there is often none to very little data loss. Managed Services Don't know why SemiAnalysis has such strong shade on "Managed Services". It's a twist of the facts. IREN has no managed services but Mirantis does. Mirantis provides managed services with 99.99% uptime SLA (15) to customers like Adobe, Cornell University, PayPal, MetLife Insurance, Royal Bank of Canada (14). With the Mirantis acquisition closed, I don't know why SemiAnalysis says that IREN advertisement of managed services is misleading, NBIS is advertising the results of their acquisitions on full blast. Conclusion @SemiAnalysis_ gives a master piece lesson on how to write FUD pieces without straight up lying but cleverly crafting context. I unravel their knots for you.

$IREN: Addressing the Semi-Analysis FUD on Reliability What's the Issue? I heard of these issue in August and the only customers that $IREN has had long enough to report "multi-day power outages, network upgrades, storage failures, air quality controls leading to link flaps, XIDs everywhere" are FireworksAI and TogetherAI at Prince George. Mackenzie is undergoing GPU commissioning as we speak (1) so I don't know why @SemiAnalysis_ mentions Mackenzie having customer problems already when there are no customers there yet. What @dylan522p says is true but convinently omitting critical industry context makes this a cleverly disguised Semi-Hit-Piece. I had heard that TogetherAI filed this complaint to IREN and Dylan was able to obtain the information from TogetherAI's CEO. I was also told that Dylan is personal friends with TogetherAI's CEO and are from the same region in India and speak the same Indian dialect which is how Dylan is able to get the info. I will show that these are industry wide problems and the affected customers renewed despite these issues. First Industry Fact First, the most important fact that @SemiAnalysis_ conveniently leaves out is that on August 27, 2026, it was confirmed that “Together AI and Fireworks AI have both renewed and expanded” (2). Despite all these problems, why did TogetherAI and FireworksAI renew? Let's look at the rest of the industry facts. Second Industry Fact Second, I will address the most alarming sounding issue: multi-day power outages. Down-time happens to everyone in the industry. @SemiAnalysis_ Platinum Cloud, $NBIS, had a outrage spanning August 19-August 20, 2026 (3). It's not narrated by SemiAnalysis in hit piece voice but the results are similar: "a storm-related event affected the building of the data center facility hosting the us-central1 region, taking the facility’s building management system offline and shutting down the chilled-water cooling loop. With active cooling lost, the data halls overheated within roughly two hours, and servers, network switches and rack power systems began shutting down on thermal protection. This caused a region-wide disruption of us-central1 services: GPU and CPU compute, external network connectivity for virtual machines, object and block storage, Managed Kubernetes, Token Factory, managed platform services, and the regional API and console endpoints" (3). "Storm-related event" makes NBIS outrage explained. SemiAnalysis's biggest bias is that they conveniently leave out the reason for IREN's outraged: BC Hydro's grid had a brown out. That's why TogetherAI renewed, not because they were okay with unexplained, random multi-day power outrages. At the time, IREN did not have enough backup generators in BC since generators produce significantly more pollution than a hydro powered grid and are hard to obtain licenses for. By March, IREN was able to obtain license to put generators for all of PG and multi-day power outrages will no longer be a problem. IREN has designed Childress and SW1 with full N+1 backup power since MSFT which requires a higher uptime which is why SemiAnalysis says those sites "look much better". Third Industry Fact Third, let's address the rest accusations: "network upgrades, storage failures, air quality controls leading to link flaps, XIDs everywhere". Using fancy technical terms without explaining them is a hallmark of a hit piece, so I'll give definitions. Link flaps means a network connection is going in and out. XID is an error code logged and reported by Nvidia GPUs. Link flaps are common: SemiAnalysis Platinum Cloud $CRWV (4), Silver Cloud Lambda (5), $MSFT (6) all had 2hr15m (4), 5hr14m (5), 4h57m (6) network out rages due to link flaps this year. Why? Because when optical networking links and transceivers collect dust particles the connection can be interrupted. SemiAnalysis conveniently uses "air quality controls" to make it sound even more scary. "XIDs everywhere" sounds scary too but they have a much larger frequency on GPUs than other compute types. $META had a paper showing that they had 419 XIDs in 54 days on their infrastructure. No other AI Cloud is this transparent about reporting problems but $META blog is one of the most honest tech blogs out there bcause META culture is "Move Fast and Break Things". Network upgrades or other control/management plane upgrades do cause problem from time to time. AWS had a DynamoDB DNS defect (8) and GCP had a policy update trigger crashes globally (9). Storage outrage is the most outrageous accusation as those are the most common across both GPU and CPU Clouds. In a completely separate incident from the storm-incident, NBIS had a storage outrage in 2025: "we had an outage of the S3 service in the eu-north1 region triggered by a steady increase of migration traffic accompanied by unexpected behavior of the program" (10). Storage downtime are not typically reported these days but in the past generation, they use to more commonly reported by GCP (11), Azure (12), AWS (13). Storage is check pointed and there is often none to very little data loss. Managed Services Don't know why SemiAnalysis has such strong shade on "Managed Services". It's a twist of the facts. IREN has no managed services but Mirantis does. Mirantis provides managed services with 99.99% uptime SLA (15) to customers like Adobe, Cornell University, PayPal, MetLife Insurance, Royal Bank of Canada (14). With the Mirantis acquisition closed, I don't know why SemiAnalysis says that IREN advertisement of managed services is misleading, NBIS is advertising the results of their acquisitions on full blast. Conclusion @SemiAnalysis_ gives a master piece lesson on how to write FUD pieces without straight up lying but cleverly crafting context. I unravel their knots for you.