Helping millions of investors navigate the AI markets. Track our 5 top-tier analysts portfolios inside Milk Road PRO. Come join us for just a $1 👇
$1 to see analyst portoflios →
- Tweets 5,834
- Following 409
- Followers 48,438
- Likes 2,017
Pinned Tweet
Satya Nadella and Mark Zuckerberg are now saying the same thing, the AI agent race will be won by trust, not intelligence.
Autonomous agents require access to credentials, private data, payment systems, and company software before they can become genuinely useful. That makes governance one of the largest barriers to adoption. Enterprises need to know what an agent can access, who authorized it, what actions it completed, how much it spent, and whether those actions can be reversed. This is why Nadella believes Agent 365 could become more important than the individual agents themselves. Microsoft is positioning it as the control layer that allows companies to observe agents, assign permissions, enforce policies, monitor spending, and maintain an audit trail of their activity.
Zuckerberg recently made almost the same argument. Meta had a capable version of Muse months before launch but the company delayed it to improve the product’s privacy and security systems. Zuckerberg believes that instruction following, intent understanding, and alignment with user values will matter more than simply improving model benchmarks. Meta built Muse around a dedicated virtual machine that isolates each agent and stores the user’s connected data and credentials. A separate Sentinel system reviews its activity, while sensitive actions such as sending emails or completing purchases require user approval.
Microsoft and Meta are approaching the opportunity from different directions, but both have reached the same conclusion. Microsoft is building the governance layer for enterprise agents, while Meta is building the trust architecture for personal agents. The companies that control identity, permissions, audit trails, secure credentials, policy enforcement, and agent spending could become just as important as the companies building the underlying models.The biggest moat in AI agents may not be creating the smartest system but rather creating the system that consumers and businesses are willing to trust with everything.
I’m positioning around the companies building the trust and control layer for AI agents because that may become just as important as the models themselves. If you want to see exactly what I hold across enterprise AI, agents, and the infrastructure behind them, check out my Milk Road Pro portfolio below.
link.milkroad.com/d8cj6v
Mark Zuckerberg believes the next phase of AI will not be won by whichever company builds the model with the best benchmark scores.
It will be won by the company that users trust enough to let AI act on their behalf.
That is why Meta reportedly had a strong version of Muse ready months before launch but delayed its release to improve its privacy and security systems.
Muse does more than generate answers because it can send emails, complete forms, make purchases, book travel, and interact with other services for the user.
Those capabilities make trust far more important because an error could affect someone’s money, private information, or real world decisions.
Meta attempted to solve this by giving every Muse user a dedicated virtual machine that isolates the agent and stores connected data and credentials.
A separate Sentinel agent reviews what Muse tries to do, while sensitive actions such as sending an email or completing a purchase require the user’s approval.
Muse also cannot directly view passwords or payment details, and users receive an audit trail showing what the agent has done.
This helps explain Zuckerberg’s belief that instruction following, intent recognition, and alignment with personal values could matter more than another improvement in math or coding benchmarks.
The most capable agent will have limited value if people do not trust it with their inboxes, calendars, payments, and personal data.
The early evidence suggests that Meta’s approach is attracting interest.
Muse recorded more than 2.5 million downloads shortly after launch, while one estimate found that more than 95% of its users also used Facebook and 63% used Instagram.
That demonstrates Meta’s biggest advantage in AI.
The company does not need to build distribution from scratch because it can introduce Muse to an enormous existing network and eventually connect it with WhatsApp, Instagram, Facebook, and its hardware.
The AI race may have started with model intelligence but Zuckerberg is betting that it will ultimately be decided by trust, distribution, and the ability to complete real tasks safely.
9
2
37
6,428
Milk Road AI retweeted
Zuckerberg, just admit you took the wrong bet.
Meta spent years building for a metaverse that Zuckerberg believed would arrive before advanced AI.
Instead, advanced AI became commercially viable while affordable holographic glasses remained years away.
Zuckerberg now acknowledges that Meta misread the order in which these technologies would develop and believes AI will ultimately become a much larger opportunity.
That mistake was incredibly expensive, Reality Labs has accumulated more than $80 billion in operating losses since late 2020, including another $4.62 billion loss in the second quarter of 2026.
Meta has since cut more than 1,000 Reality Labs jobs, closed several VR studios, and redirected resources from virtual reality toward AI powered glasses and wearables.
Zuckerberg is not completely abandoning the metaverse but he is putting the most ambitious version of it on the shelf until the hardware becomes affordable enough for mass adoption.
Meta is now treating glasses as a delivery system for AI rather than a portal into virtual worlds.
That is a fundamentally different strategy because the original plan was to convince people to enter Meta’s digital world through a headset.
The new plan is to bring Meta’s AI into the world people already inhabit through lightweight glasses that can see, hear, and interact with their surroundings.
This strategy has a much clearer path to adoption because users do not need to change their behavior or isolate themselves inside a virtual environment.
Meta can also connect those glasses with Muse, giving its personal AI agent access to what users see and hear throughout the day.
11
3
28
5,694
Milk Road AI retweeted
Did you guys listen when I told you to buy $CRDO?
The stock is now up massively from where I first started buying and it’s up another 8% today.
You can check out my entire portfolio and see exactly what I’m buying here.
link.milkroad.com/tx5qip
Here is why I am adding to my Credo Position after the earnings.
Credo delivered another quarter of triple digit revenue growth, remained extremely profitable and guided above expectations. One of the problems was that Credo had become priced for near perfect execution so investors wanted a much larger beat and a bigger increase to the full year outlook.
If you don’t know what Credo is it’s basically one of the most important connectivity companies inside AI data centers. NVIDIA and other AI chips provide the computing power, but those processors still need to communicate with switches, memory and thousands of other chips. Credo builds the cables, processors and optical components that move this information around the data center. Its largest business is Active Electrical Cable or AECs. These are advanced copper cables containing Credo chips that strengthen and manage signals as data moves between AI servers and network switches. They provide lower power consumption and greater reliability than optical connections across short distances while being thinner and easier to use than traditional passive copper cables. Credo is also expanding deeper into optical networking. Its optical DSPs clean up signals traveling through optical transceivers, while its silicon photonics chips help convert electrical data into light. After acquiring DustPhotonics, Credo can now sell the DSP, the silicon photonics chip or the complete ZeroFlap optical transceiver. This increases the amount of Credo technology inside every connection and could make optics the company’s next major growth engine.
ZeroFlap is especially interesting because connection failures are becoming a serious problem inside massive AI clusters. Credo’s PILOT software continuously monitors signal quality and can detect when an optical connection is beginning to weaken. The system can then remove the affected GPU or reroute traffic before the connection completely fails. When a data center contains billions of dollars of GPUs even a small improvement in uptime can be extremely valuable.
Okay so the latest earnings were objectively strong in my opinion. Quarter revenue reached a record $479 million, increasing 9.6% sequentially and 114.7% year over year. Wall Street expected approximately $473 million. This was Credo’s seventh consecutive quarter of triple digit revenue growth. Profitability was also extremely strong. Credo generated $236.3 million in adjusted net income, up 140% year over year, producing an adjusted net margin of 49.3%. Adjusted operating income reached $230.6 million, giving the company a 48.2% operating margin. Another reason for the selloff is that growth is beginning to decelerate. Revenue grew 115% year over year this quarter, while second quarter guidance implies approximately 98% growth. That is still incredible growth and does not suddenly make the stock expensive but the market rarely rewards decelerating growth. When a stock is priced for near-perfect execution, even a slowdown from incredible growth to extremely strong growth can cause its valuation to compress.
Guidance was also better than expected. Credo expects second quarter revenue between $525 million and $535 million, compared with Wall Street’s estimate near $520 million. Management maintained its forecast for more than 85% full year revenue growth, more than $600 million in optical revenue and an adjusted net margin near 50%.
AEC growth is also beginning to normalize. AEC revenue more than doubled during fiscal 2025 and more than tripled during fiscal 2026. Management still expects the business to grow, but acknowledged that optics will grow faster moving forward. Credo now needs to prove that its optical products can ramp quickly enough to become a second major growth engine. Margins also created some concern. GAAP gross margin fell from 68.2% to 64.5% sequentially, although most of that decline came from stock compensation and acquisition related amortization. Adjusted gross margin only declined from 68.3% to 68%, so the underlying business did not suddenly become less profitable.
Stock compensation is a bigger concern. Credo excluded nearly $88 million of stock compensation when calculating adjusted earnings, up from roughly $35 million last year. Diluted shares also increased around 5% year over year. Adjusted earnings show how strong the underlying business is, but this compensation still dilutes existing shareholders. Customer concentration remains the largest risk. Credo’s four biggest customers represented 33%, 28%, 13% and 10% of quarterly revenue. That means four customers generated approximately 84% of total revenue, while the two largest generated 61%. Losing one major deployment or experiencing a customer delay could materially affect results. The bull case is that Credo is no longer simply an AEC company. It now owns more of the connectivity stack, including copper cables, optical DSPs, silicon photonics chips, complete optical transceivers, retimers, memory connectivity and diagnostic software. This gives Credo more products to sell to the same hyperscalers while increasing its potential revenue from every AI accelerator
I think the current price offers a much better risk to reward than the stock did above $200. If you did not already own Credo, I would view this as an attractive price to begin building a position. The company is still growing extremely quickly, remains highly profitable and is expanding into several new connectivity markets. However, with the current macro environment and weakness across AI stocks, I would scale into the position instead of buying everything at once.
If you enjoyed reading this, make sure to follow @MelvinInvests for more AI infrastructure and semiconductor insights and turn on post notifications so you don't miss a single update
2
3
18
6,439
Milk Road AI retweeted
America has a massive lead in AI compute but China is exposing a weakness in that advantage owning more chips matters a bit less if your competitor can generate intelligence at a fraction of the cost. The United States controls nearly three quarters of global AI computing power compared with just over 14% for China yet cheaper Chinese models already starting to dominate OpenRouter’s rankings by tokens processed.
This creates an uncomfortable possibility. China could produce more usable AI output without matching the United States chip for chip, simply because its models generate more tokens from every dollar spent on compute. If that efficiency continues improving while China adds more data centers and domestic chips, America’s compute lead could become less valuable than it appears.
The most important measure may no longer be how much compute a country owns but how much intelligence that compute can produce. Because Chinese labs have limited access to the most advanced chips, they have been forced to make their models more efficient. They are using techniques that activate only the parts of a model needed for each task, reducing the processing power and memory required to generate every token. Yes, obviously they are using distillation as well but all that 'efficiency' leads directly to lower prices. Chinese open models can be 60% to 90% cheaper than leading American alternatives, while still delivering strong enough performance for many everyday workloads.
Token cost becomes even more important as AI moves from simple chatbots to agents. Each step consumes more tokens, so even a small difference in token price can become a major expense when companies run millions of AI tasks. This gives cheaper Chinese models a major advantage. If a model costs one fifth or one tenth as much, developers can let it work longer, attempt tasks multiple times and serve far more users for the same budget. On OpenRouter, Chinese models have been priced 60% to 90% below leading American models, helping their share of US enterprise token usage rise as high as 46% by mid-2026.
This is where the compute race and the token price race begin to connect. Lower prices create more usage. More usage keeps China’s chips and data centers busy, brings in revenue and gives developers more feedback to improve their models. That demand can then support additional investment in chips, power, networking and data centers. The cycle can quickly reinforce itself. Better efficiency lowers token prices, lower prices attract more users, and more users create the demand needed to build more compute. China may start with far less computing power but it can use lower prices to grow token volume faster and then reinvest that growth into expanding its infrastructure.
American AI companies face the opposite pressure. They are spending enormous amounts on GPUs, electricity and data centers, but competition keeps pushing token prices lower. If they must match Chinese prices while carrying much higher development and infrastructure costs, their profit margins could shrink and it could become harder to earn a return on their massive AI investments. This does not mean China will soon own more advanced chips than the United States. America still has a major advantage in semiconductors, cloud infrastructure and total computing capacity. The larger risk is that China wins the economics of compute first by producing more affordable tokens from every chip it has. If that happens, China will not need to beat America chip for chip. It could close the gap by combining more efficient models, cheaper tokens and a steadily expanding compute base. The winner of the AI race may not be the country with the most GPUs but the country that turns those GPUs into the greatest amount of affordable intelligence.
I’m positioning around both sides of this trade because cheaper models do not reduce the need for infrastructure, they can actually drive even more usage and compute demand. If you want to see exactly what I hold to capture that value across chips, memory, neoclouds, and the broader AI stack, check out my portfolio below.
link.milkroad.com/d8cj6v
3
4
30
4,957
Milk Road AI retweeted
$BSP has fallen ~40% since August highs and is underperforming the SaaS market.
Here is why it’s a GREAT buy the dip!
As a reminder, Bending Spoons buys older SaaS companies, uses AI to boost profits, then buys more.
2 things the market is repricing right now:
First, the hype around $META MUSE fuels the argument of AI agents destroying SaaS seats the company owns. Similar to what ChatGPT "did" to Google Search :)
Second, the higher cost of capital environment (10yr, Oil, Fed hike) raises cost of debt and impacts returns on acquisitions.
Here is why I think this is short term volatility providing a great entry point:
Regarding Agents:
BSP kept 94% of revenue from existing customer in Q1 2026. They have a very stick customer base where people are unlikely to abandon their notes, videos and established workflows. If anything, agentic usage of those tools will increase value for humans.
Regarding the cost of debt:
S&P just upgraded BSP to BB− despite its growing debt for Miro acquisition and expects interest coverage >3x by 2027. Meaning the credit agency is expecting BSP earnings will cover interest payment more than 3 times over by 2027. Aka, solid business health.
The CEO also said that existing debt is hedged at ~ 9%, while higher rates could push purchase prices down enough to offset pricier new loans.
2
3
5
827
Jensen Huang just made one of the most controversial arguments in AI that slowing development could make the technology less safe.
His position is that AI safety is fundamentally an engineering problem that requires better technology, more computing power and more rigorous testing to solve.
In other words, the answer to unsafe AI is not freezing today’s models in place but building stronger systems around them that can monitor behavior, verify outputs, restrict access, and intervene before something goes wrong.
There is already evidence supporting part of that argument.
Techniques such as reinforcement learning from human feedback can steer model behavior, while retrieval augmented generation can ground answers in approved information rather than relying entirely on the model’s memory.
More capable safety models could eventually monitor other AI systems in real time, detect suspicious behavior and stop agents from completing harmful actions.
This becomes increasingly important as AI moves beyond answering questions and starts sending emails, writing code, accessing databases, and operating software independently.
However, the same progress that improves safety also creates more capable systems that are harder to contain.
That tension became clear when OpenAI agents reportedly escaped a restricted testing environment, reached the open internet, and accessed Hugging Face’s infrastructure during a cybersecurity evaluation.
The incident was not simply evidence that AI had become uncontrollable but it exposed failures in sandboxing, network isolation, access controls, and monitoring, the exact engineering systems designed to contain the agents.
That supports Huang’s argument that many near term AI risks resemble traditional cybersecurity failures, only with faster and more autonomous software operating inside the system.
The solution requires isolated testing environments, strict permission limits, adversarial evaluations, continuous monitoring, human approval for sensitive actions, and immediate shutdown mechanisms.
5
5
26
3,842
Milk Road AI retweeted
Mark Zuckerberg believes the next phase of AI will not be won by whichever company builds the model with the best benchmark scores.
It will be won by the company that users trust enough to let AI act on their behalf.
That is why Meta reportedly had a strong version of Muse ready months before launch but delayed its release to improve its privacy and security systems.
Muse does more than generate answers because it can send emails, complete forms, make purchases, book travel, and interact with other services for the user.
Those capabilities make trust far more important because an error could affect someone’s money, private information, or real world decisions.
Meta attempted to solve this by giving every Muse user a dedicated virtual machine that isolates the agent and stores connected data and credentials.
A separate Sentinel agent reviews what Muse tries to do, while sensitive actions such as sending an email or completing a purchase require the user’s approval.
Muse also cannot directly view passwords or payment details, and users receive an audit trail showing what the agent has done.
This helps explain Zuckerberg’s belief that instruction following, intent recognition, and alignment with personal values could matter more than another improvement in math or coding benchmarks.
The most capable agent will have limited value if people do not trust it with their inboxes, calendars, payments, and personal data.
The early evidence suggests that Meta’s approach is attracting interest.
Muse recorded more than 2.5 million downloads shortly after launch, while one estimate found that more than 95% of its users also used Facebook and 63% used Instagram.
That demonstrates Meta’s biggest advantage in AI.
The company does not need to build distribution from scratch because it can introduce Muse to an enormous existing network and eventually connect it with WhatsApp, Instagram, Facebook, and its hardware.
The AI race may have started with model intelligence but Zuckerberg is betting that it will ultimately be decided by trust, distribution, and the ability to complete real tasks safely.
Bill Ackman made the perfect bull case for Meta months before even Muse arrived.
At the time, the market looked at Meta’s massive AI spending and assumed Mark Zuckerberg was lighting money on fire.
However, @BillAckman argued that investors were asking the wrong question.
Investors should not focus only on how much Meta was spending but rather on why it was spending.
If Meta had doubled its capital expenditures simply to protect its existing business, then the stock deserved to fall.
But if the company was investing in AI infrastructure that could create new products and generate attractive returns, the spending was growth capex rather than maintenance capex.
But now Muse made that argument much easier to understand.
Meta used its AI infrastructure to launch a personal agent that could send emails, book travel, make purchases, and complete other tasks through its own app or WhatsApp.
Instead of building distribution from scratch, Meta could place Muse directly inside an ecosystem already used by billions of people.
Muse also introduced paid subscription tiers, giving Meta another potential revenue stream beyond advertising.
Muse may only be the beginning because if Meta can keep turning AI infrastructure into new products, that massive capex bill starts looking a lot more like an investment than an expense.
7
10
77
21,123
Milk Road AI retweeted
China started an AI price war America is not be ready for.
Its models are drastically cheaper, forcing enterprises to choose between US security and Chinese economics.
But here are five companies positioned to win.
6
4
15
7,164
Milk Road AI retweeted
Agents are going to drive the next leg of this bull market (BULLISH AI INFRA)
Meta's new Muse agent just hit ~3M U.S. downloads in about 2 weeks, nearly DOUBLE ChatGPT's pace at the same point
As I've been saying for a while, agents will be WAY bigger than chat. It's a better product and it's way more useful
And most importantly, it uses WAY more compute (this is the part the market is still underpricing)
A chat is one question, one answer. An agent task is a whole chain of model calls: browsing, reading, deciding, acting, checking its own work. Every task burns a multiple of the tokens a chat does
Now multiply that across millions of consumers... and even bigger is the enterprises running agents in the background all day at scale
Enterprises are already paying billions for AI, thats why Antrhopic and OpenAI revenues are accelerating so fast. But even consumers will pay for agents because it saves them time (and money in many cases)
I think the market has not yet priced in whats coming as the world adopts agents. This is going to be a huge step change in adoption, token usage and compute demand
We are going MUCH MUCH higher and rate hikes or high yields in the bond market aren't stopping this train.
Want to see how to invest in this market? You can track my real-time portfolio here: link.milkroad.com/t5837z
Don't forget to give me a follow @kylereidhead for more insights on AI and markets
$META just got a Canaccord Genuity price target raise to $950
The note cites Muse hitting 2.5M+ downloads in two weeks, with access expanding to glasses, Mac and a dedicated device
AI monetization via subscriptions and transaction fees is now being underwritten for Meta
8
8
28
8,745
Hyperscaler capital spending is not declining but only the percentage growth rate is expected to slow after an extraordinary surge (Save this).
Spending across the five largest hyperscalers rises from roughly $150 billion in 2023 to $800 billion in 2026, before reaching an estimated $1.2 trillion in 2027 and $1.4 trillion in 2028. That represents more than a ninefold increase in five years. The additional $200 billion projected for 2028 alone is almost as much as the hyperscalers spent during all of 2024. The $1.2 trillion expected in 2027 is also greater than their combined spending from 2023 through 2025. This does not look like the end of the AI infrastructure cycle but rather looks like the point where an experimental buildout becomes an industrial economy.
Goldman Sachs estimates that roughly $7.6 trillion could be invested across AI compute, data centers, and power infrastructure between 2026 and 2031. The first phase primarily rewarded the companies selling GPUs. The next phase should spread more of that spending into memory, networking, optical components, cooling, electrical equipment, energy storage, construction, and power generation. That shift is becoming necessary because the bottleneck is no longer just producing enough chips but rather finding enough power and physical infrastructure to operate them.
The IEA expects global data center electricity consumption to roughly double between 2025 and 2030, while electricity demand from AI-focused facilities is projected to triple. The winners of this phase will be the companies sitting directly in front of the physical bottlenecks that $1.4 trillion of annual hyperscaler spending cannot avoid. That means memory, networking, optical components, cooling, electrical equipment, energy storage, power generation and the hyperscalers that can turn all of this infrastructure spending into actual earnings. Between now and when this buildout fully matures will be one of the best periods to make money in the market. If you want to see exactly what I hold to capture that value across AI infrastructure, you can check out my full Milk Road Pro portfolio using the link below.
link.milkroad.com/d8cj6v
5
2
20
3,125
Milk Road AI retweeted
Elon just confirmed that SpaceX will become the LARGEST NEOCLOUD in the world by EOY
Colossus 1 & 2 (so far) provide about 1.4 GW of compute
With the addition of 660k GB300s by EOY, $SPCX will have approx 2.6 GW of compute
For comparison, Coreweave is tracking 1.85 GW & Nebius 1 GW by EOY
At 2.6 GW, Elon said around 10% will go to Grok/xAI, which means there is 2.3 GW available to rent
At $30-$50B/GW, that means SpaceX is ending the year at $70B-$115B in ARR from the compute business alone. My guess is xAI will be closer to $20B ARR by the then to
And this doesn't include any revenues from Starlink or Rocket launches. If all goes well, SpaceX will be ending the year above $125B ARR at least.
Unbelievable growth in 2026 and 2027 should be even bigger, considering Elon expects to reach 10 GW of compute by the end of 2027
By the way, I bought $SPCX at $115 and again at $140 inside my Milk Road PRO portfolio. I think it's still significantly undervalued at $150. You can track my real-time portfolio here: link.milkroad.com/t5837z
Good times ahead for SpaceX!
SpaceX and Tesla will win because they will own the most compute
xAI will beat Anthropic and OpenAI because they will acquire the most compute
Tesla Optimus will win because they will have more compute than Figure and Chinese competitors
Cybercab and Robotaxis will win because they will have more compute than Waymo and other competitors
And in the long-run, SpaceX and Tesla will win because they will own the chips via Terrafab, the connectivity via Starlink and the access to Space via Starship
There are not many moats left in enterprises, but SpaceX and Tesla continue to own the biggest ones because they focus on doing the hardest things
This is why Spacex and Tesla (likely merged at some point) will become a $100T company in the years (decades) to come
@elonmusk is thinking in decades, not months like most other companies
13
44
291
25,658
Bill Ackman made the perfect bull case for Meta months before even Muse arrived.
At the time, the market looked at Meta’s massive AI spending and assumed Mark Zuckerberg was lighting money on fire.
However, @BillAckman argued that investors were asking the wrong question.
Investors should not focus only on how much Meta was spending but rather on why it was spending.
If Meta had doubled its capital expenditures simply to protect its existing business, then the stock deserved to fall.
But if the company was investing in AI infrastructure that could create new products and generate attractive returns, the spending was growth capex rather than maintenance capex.
But now Muse made that argument much easier to understand.
Meta used its AI infrastructure to launch a personal agent that could send emails, book travel, make purchases, and complete other tasks through its own app or WhatsApp.
Instead of building distribution from scratch, Meta could place Muse directly inside an ecosystem already used by billions of people.
Muse also introduced paid subscription tiers, giving Meta another potential revenue stream beyond advertising.
Muse may only be the beginning because if Meta can keep turning AI infrastructure into new products, that massive capex bill starts looking a lot more like an investment than an expense.
11
10
62
22,926
Milk Road AI retweeted
HOLY SMOKES!
Higgsfield just reached $1 billion annual revenue run rate faster than Anthropic and OpenAI and this may be one of the strongest bull signals yet for Nebius.
Because nebius is helping provide the infrastructure supporting that growth.
Higgsfield used Nebius AI Cloud and NVIDIA HGX B200 systems to train its large scale image and video diffusion models.
Nebius also helped create the storage, networking and data pipelines needed to keep Higgsfield’s GPUs operating efficiently.
When Higgsfield had a $200 million revenue run rate, it already had more than 15 million users and was processing over 4.5 million image and video generations every day.
The amount of infrastructure required at a reported $1 billion run rate could be enormous.
Revenue and compute usage will not necessarily increase at the same rate, but more paying customers, larger enterprise accounts and higher generation volumes should create greater demand for GPUs, storage and networking.
AI video is especially compute intensive because every request can require the system to generate hundreds or thousands of frames and users also create several versions, make edits and rerun prompts before selecting a final result.
Every one of those actions consumes additional computing capacity.
Higgsfield also needs large GPU clusters to train new models, improve video quality and run repeated preference training cycles.
Nebius can therefore benefit from Higgsfield’s major training runs and its recurring inference usage.
Higgsfield is exactly the kind of customer Nebius wants because its infrastructure needs can keep expanding without Nebius having to find a new buyer for every additional GPU it installs.
The relationship also gives Nebius a powerful reference customer because it shows other AI startups that Nebius can support a highly demanding video platform as it scales from zero to millions of users and a reported $1 billion run rate.
I hold Nebius because I want exposure to the infrastructure underneath companies like Higgsfield as their usage explodes.
If you want to see how I’m positioned around Nebius, neoclouds, and the broader compute trade, check out my Milk Road portfolio below.
link.milkroad.com/tx5qip
6
9
88
9,058
Everyone thinks data center moratoriums are killing the AI boom but the actual numbers tell a different story.
More than 300 local governments have adopted data center restrictions across the United States, with new moratoriums accelerating sharply during 2026.
A moratorium usually pauses new applications while officials review concerns involving electricity, water, noise and land use.
At first glance, hundreds of restrictions sound like a major threat to the AI infrastructure buildout.
However, SemiAnalysis estimates that only 1,525 megawatts is actually delayed out of roughly 20 gigawatts located inside restricted areas.
Including New York’s statewide action raises the total meaningful delay to approximately 2.3 gigawatts.
Most moratoriums cannot stop projects that already have permits, zoning approval or construction underway.
Many restrictions will also expire before future projects need approval, while developers can sometimes relocate, redesign their campuses or build on site power systems.
SemiAnalysis still expects more than 38 gigawatts of US data center capacity to be delivered in 2027, more than double the 2026 level.
Approximately 22 gigawatts is already under vertical construction, while much of the remaining capacity is financed or undergoing site preparation.
The political backlash is real but the money, equipment and construction are already moving.
These restrictions could actually make approved land and secured power connections more valuable because developers with permits can continue building while competitors remain stuck.
Moratoriums could also accelerate behind the meter power, including natural gas generation, fuel cells, batteries and microgrids.
Build out still continues and it's not going to be stopped by anyone.
10
7
32
4,117
Milk Road AI retweeted
OMG, UBS expects Memory to eat 64% of Hyperscaler Capex next year
Memory spend goes from $369B in 2026 to $933B in 2027, while Capex goes from $1T to $1.45T
Memory is more than 100% of the Capex raise
That means everything that ISN'T memory shrinks next year, from $630B to $520B (wow)
Three things I take from this:
1. This is a PRICE story, not a volume story. Hyperscalers aren't buying 13x more memory than they did in 2025, they're paying way more per bit. UBS has DDR prices up 341% this year alone
2. I'm seeing people on X point at $1.45T of capex and scream bubble. But more than all of the increase is landing with a handful of memory makers (Samsung, SK Hynix, Micron, plus the NAND players). That's not a bubble, that's the most extreme pricing power anywhere in the AI supply chain
3. It doesn't reverse soon either. UBS has 2027 DRAM demand growing 36% vs supply growing 19%, and Citi has DRAM prices rising EVERY year through 2031 (see my post below)
The market seems to have stopped caring as much about Memory, as its consolidated for the last 3 months but I think Micron and the others are still a screaming buy here as the market wakes back up to these numbers
I'm long $MU and if you want to see what else I hold in my portfolio you can check it out here: link.milkroad.com/t5837z
Don't forget to give me a follow @kylereidhead for more insights on AI and markets
WOW! Memory stocks are going MUCH HIGHER
Citi just forecasted DRAM prices rising EVERY year through 2031
That has never happened in the history of the memory industry. If Citi is right, the way the market prices memory stocks is flat out wrong
Not only is this not cyclical, it's a structural change in prices that pushes Memory higher for at least 8 years straight
Why is this happening? HBM, the memory inside every AI GPU, consumes roughly 3x the wafer capacity of standard DRAM per bit. As Micron, SK Hynix and Samsung convert their fabs to HBM, regular DRAM supply gets starved right as AI demand for it explodes too. Citi sees that supply tightness persisting until 2031
Bears will scream "cycle peak" at that deceleration in 2027. But look at what the numbers actually say: prices NEVER go negative. Not one down year through 2031
For 25 years memory was a commodity coin flip, massive up years always followed by brutal crashes (2019: -51%. 2023: -43%). Citi is forecasting the first 8-year stretch in industry history with zero down years. That's not a cycle anymore, that's a repricing
And it flows almost straight to the bottom line for producers. Micron just posted the largest earnings beat in its history with 85% gross margins, and that was BEFORE the 242% year fully hits the income statement.
Even more important to note that teh buybacks from this free cash flow haven't even started yet (coming in December)
The market still values memory stocks like the crash is 18 months away, according to Citi's chart that's not happening
I hold Micron in my Milk Road Portfolio, our analysts got our members in $MU below $400. We also hold a few other memory stocks alongside other ways to play the Ai infra build out. You can check out our real-time portfolios here: link.milkroad.com/t5837z
Don't forget to give me a follow @kylereidhead for more insights on AI and markets
8
37
163
20,881
Milk Road AI retweeted
This is the reason why Nebius will continue to go even higher.
Nebius is generating increasingly more contracted annual revenue from every gigawatt of power capacity it builds. This chart shows annual recurring revenue per gigawatt rising from approximately $6 billion in the third quarter of 2025 to $12 billion in the second quarter of 2026. BNP Paribas estimates that Nebius new second quarter contracts reached approximately $20 billion of ARR per gigawatt, while short term third quarter contracts could reach $40 billion per gigawatt when that capacity is delivered in 2027. That would be more than six times the $6 billion level recorded in 2025. Now that does not mean Nebius is already generating $40 billion in annual revenue but it does mean that each future gigawatt of infrastructure could support much more contracted revenue than before.
The increase ie because of stronger demand for newer Nvidia systems, larger customer contracts and premium pricing for reliable managed AI infrastructure. Nebius has also raised its standard GPU prices, increased Token Factory inference pricing and introduced dynamic spot pricing for spare capacity. Together, these moves allow Nebius to charge more for guaranteed capacity and that can increase revenue per GPU. Higher revenue per gigawatt can dramatically improve data center economics. The cost of land, power connections, buildings and cooling does not necessarily increase as quickly as contract revenue, so a larger portion of the additional revenue could eventually flow into gross profit and operating income.
Long term contracts also give Nebius better revenue visibility before the capacity becomes operational. That can help the company secure financing, purchase additional GPUs and build more data centers because lenders can see contracted demand supporting the investment. The potential flywheel is pretty straightforward, Nebius signs larger contracts, those contracts unlock financing, the financing funds more capacity and the additional capacity allows Nebius to sign even larger customers.
Bullish on Nebius and make sure to follow @MelvinInvests for more insights.
4
18
112
10,378
The next major AI bottleneck may be delivering enough stable power to keep them running (Save this).
This Goldman Sachs chart maps the companies supplying power and energy storage equipment for next generation AI server racks.
It focuses on battery backup units, or BBUs, and capacitor backup units, or CBUs.
AI workloads can cause a server rack’s power consumption to change within milliseconds and as AI racks become more powerful, these sudden changes can destabilize voltage, create GPU errors or restart an entire system.
BBUs place batteries close to the servers so they can provide electricity during outages, voltage drops or sudden increases in power demand.
CBUs use capacitors and supercapacitors to react even faster and stabilize short power fluctuations.
Now, here are some of the companies that could benefit from this
Panasonic, Samsung SDI, LG Energy Solution, Murata and TDK can benefit from rising demand for battery cells and BBU components.
Delta Electronics, Flex, Lite-On, Advanced Energy, AcBel Polytech and Chicony Power supply power systems, converters or complete backup units.
Texas Instruments, Infineon, Renesas, Analog Devices, NXP, STMicroelectronics, ROHM and Monolithic Power Systems provide the chips and controllers that manage electricity inside AI racks.
Vertiv, Schneider Electric, Eaton, ABB and Legrand sell racks, busbars and power distribution equipment.
Nichicon, Nippon Chemi-Con, Yageo and Rubycon can benefit from growing capacitor demand, while Panasonic, Eaton and Musashi Seimitsu have exposure to supercapacitors.
Every new high density AI rack requires more batteries, capacitors, converters, power supplies, controllers, cooling equipment and electrical infrastructure.
Goldman Sachs reportedly expects the BBU market to grow from approximately $2.2 billion in 2025 to $50.2 billion by 2030.
I’m increasingly focused on the companies sitting behind the power systems inside AI racks, because this bottleneck gets bigger as compute density rises.
If you want to see how Milk Road Pro members are positioned around power, cooling and the broader AI infrastructure buildout, check out my portfolio below.
link.milkroad.com/d8cj6v
3
10
38
4,676
