Software vs Hardware: The CAPEX Risk

The AI trade so far has paid the companies selling infrastructure. But hardware revenue depends on the rate of change of customer spending, and that rate is forecast to roll over: combined hyperscaler capex has roughly tripled from $256 billion in 2024 toward ~$785 billion in 2026, with Moody’s projecting nearly $1 trillion for 2027 implying growth decelerates from ~70% to roughly 27%.

Analysts have flagged the timing of this slowdown as a direct risk to infrastructure valuations, and the strain is showing: capex now approaches or exceeds operating cash flow at some operators (vs. a ~40% ten-year average), funding has shifted toward debt, and Moody’s has warned of creditworthiness reassessments if AI profits lag. Power constraints and shareholder ROI demands make the deceleration a question of when, not if.

Software inverts the exposure. Subscription revenue rides on the cumulative stock of deployed compute and workflows which keeps growing even after spending decelerates.

Deployed GPUs become cheap, creating even more inference capacity: the raw input cost of AI software. Falling compute costs (Alphabet cut Gemini serving costs 78% in 2025) squeeze hardware pricing power while expanding software gross margins.

By seeing how software multiples have lagged the infrastructure trade badly, AI monetization is arguably not yet fully priced in at the multiples we see today.

This is the exact reason whenever hardware stocks like AMD, AVGO, e.t.c fall software stocks always seem to inverse and go up.

Big money is trying to squeeze every last drop of profit out of hardware before the inevitable shift into software companies, making software companies a very attractive investment to those with larger time horizons, which is the opposite of big money trading firms who need results every quarter to impress investors/shareholders leading to the cheap valuations we see across the software space.

Software Names Poised to Win + Price Targets

Now there are many names in this space that are truly great companies with very healthy balance/income/cashflow statements trading at attractive valuations.

The names everyone knows like Microsoft, Meta, ServiceNow all have great valuations and would realistically all succeed when CAPEX dwindles away. But we at Aria Research are really focused on 3 big names here. Microsoft (which we have written about before and will therefore be omitted here - see former article), Salesforce (CRM) and a more speculative but interesting prospect in the space UiPath (PATH).

Here I will give our analysis on the direct head to head of two of our 3 favorites in the space with our price targets and reasoning behind them…

Salesforce ( ) vs UiPath ( )

Both names sit at the application layer, but they express the thesis differently: Salesforce is proven agentic monetization at scale trading at a depressed multiple.

UiPath is a cheap, newly profitable platform with unpriced optionality on agent orchestration.

Read more

Muse Spark 1.1

Meta’s newest most powerful API model is Muse Spark 1.1. On their first day of public access, it outperformed some of the leading frontier models on agentic workflows, such as Anthropic’s Opus 4.8 on MAX.

Source: Meta.AI

From the analysis above, Muse Spark 1.1 is the current leader in agent benchmarks in 4/6 categories. Where it fails to beat Claude and OpenAI is in hardcore coding (SWE-Bench, DeepSWE) and Multimodal categories, which test software engineering and the ability to analyze, reason across, and generate different formats.

For Meta, this is a generational breakthrough. For years, analysts have been scorning Zuckerberg for his lacklustre AI spending, committing hundreds of billions of dollars to a buildout that simply didn’t have a clear link to revenues. Their underlying ads and social media business was more than able to fund the buildout; the question of ROI, however, was always overhanging. We believe that Muse Spark 1.1 is the first of many strides forward that will crown Meta as the leader of AI.

Meta now has a competitive frontier model and is monetizing it through token usage. This is the first time in Meta’s history that they’ve made users pay for their AI models, and it’s clear what their objective is. They are trying to offer the most efficient agentic model that can operate at the lowest cost per token. Meta will pose a serious threat to both OpenAI and Anthropic simply because of their token costs and heightened efficiency.

Source: Meta.AI

The Tokenomics:

As we’ve come to understand, the newest, most powerful GPUs are reserved specifically for training large-parameter frontier models. As newer generations of chips get released, the older ones slowly lose their value. H100 and H200 equivalents used to be the most powerful on the market and have been pushed toward less intensive use cases. Simply put, these chips no longer compare to the current best on the market, but that doesn’t mean they’ve gone entirely obsolete. These older-generation chips are still perfectly viable for inference and agentic use cases.

Source: Silicon Data

This chart from Silicon Data models the H100 term rate curves, showing a substantial rise in prices through 2026 when compared to November, 2025. This chart makes economic sense as to why Meta is interested in entering the “neocloud” business, as the assets they’ve stockpiled are appreciating. What Meta is aiming to achieve does not signify a compute oversupply, but rather a generational shift in financing compute. The vast majority of chips that are being rented out from Meta will be their old stockpile of H100 and H200’s, while maintaining Blackwell and Rubin chips for their frontier model training. Those higher-end GPU’s will be supply-constrained through 2028, as every single company on the planet requires them to be competitive. What also becomes supply-constrained will be the older GPUs that are now obsolete for frontier training. As we know, leading-edge GPU’s like Rubin and Blackwell deliver roughly 3x tokens per megawatt versus H100, with Nvidia claiming up to 15x lower cost per token on optimized workloads. In a vacuum, the leading-edge GPUs are a no-brainer choice, but not when they are completely sold out. This AI super cycle is causing the strongest compute to be utilized specifically for training frontier models. Even Jensen Huang, while announcing that Nvidia had secured supply for ‘‘very, very robust growth,’’ admitted the company remains supply constrained.

Blackwell's headline advantages, such as FP4 compute, giant NVLink domains, and 8 TB/s bandwidth, matter most for serving massive frontier models at high interactivity. But the token volume concentrated at the frontier could be rather misleading: Beneath it sits a huge layer of everyday enterprise AI. Small fine-tuned models, embeddings, batch summarization, document-lookup pipelines, and ordinary chatbots that don’t need frontier models or frontier hardware at all. For a bandwidth-bound 70B decode workload, an H100 isn’t something you should frown at; it's actually correctly sized to operate under these parameters.

The trade-off is token efficiency; on SemiAnalysis's benchmarks, Blackwell serves tokens 4–9x cheaper than Hopper on optimized workloads. But per-token efficiency only matters if you can get the hardware, and for workloads that fit comfortably in an H100's 80GB, oftentimes the cheapest option becomes the most rational choice. Market data pulled from SemiAnalysis’s InferenceMAX platform shows that the cost per 1 million tokens for an H100 and a Blackwell chip per 1 million tokens is $0.14 and $0.03, respectively. The reason Hopper chips are even viable is the spread between token prices and token costs. API tokens for large open models retail at $0.20-$1.00 per million; SemiAnalysis's benchmarks put serving costs for these models at $0.03–0.55 per 1 million tokens, depending on the chip. At spreads like that, even hardware with several times worse token economics still runs at enormous gross margins. The efficiency gap between legacy and frontier chips decides who earns more, and not who earns.

The Hardware Bonanza:

Hardware was sold off hard over the past week due to the misconception around Meta’s announcement. Many individuals took the selling of “excess” compute as an indication of an AI oversupply. We are here to tell you that this is factually incorrect. To begin, if there were truly an oversupply of compute, then why is Meta renting 1.6 GW worth of capacity from Crusoe, committing $27 billion to Nebius, $35.2 billion to CoreWeave, and just recently having Google cap their Gemini usage? Companies that are drowning in “excess” capacity do not exhibit behaviours like this. So why are they? The answer is only found between the lines… Meta selling their older generation chips while simultaneously locking up neocloud capacity is not contradictory. They are selling yesterday’s viable compute, while locking up tomorrow’s scarce capacity. This is Meta’s way of monetizing every part of their AI business, while also funding the next round of Capex. One of Meta’s largest criticisms was their lacklustre spending with no real sign of ROI. Muse Spark 1.1 and their “neocloud” proposal refute all claims of “lack of ROI.” The FCF and narrative legitimacy they can acquire from this can help Meta raise Capex guidance without being punished for it; AKA fuelling the AI trade for even longer. We believe this will cause a snowball effect of Capex guidance being revised upward. We predict that Capex from hyperscalers could see $1.0-$1.3 trillion in 2027.

Source: Aria Research

Compute & Neoclouds:

The following section will talk thoroughly about the impacts on the compute and neocloud sectors.

Read more

As you know, at Aria Research, Intel is one of our highest-conviction companies, so it is only appropriate that we provide an in-depth description of Intel’s foundry pipeline, along with our view of its stock valuation through 2030.

We are going to upload the same Excel File we used to model Intel’s valuation for Paid Subscribers. The Excel file we created allows you to “plug and play” with numbers and scenarios, so even YOU could make your own valuations. That being said, let’s embark on the “Intel model”.

Pillar 1: Assumptions

With our valuation model, there are some assumptions we have made in order to forecast EPS. We purposely left out any incremental AI accelerator revenue over the next 5 years, as they currently hold negligble market share in this space. We assumed as well that Intel will maintain 25% market share in server CPU’s, with a TAM of $170 billion. Anything related to the consumer market is priced into the base EPS of $6.24. This articles purpose is to ONLY model the Foundry pipeline through 2030…

Source: Aria Research

The Tera-Fab Deal:

The Tera-Fab project is a $55B joint venture between Tesla, SpaceX, and xAI, unveiled by Musk in late March 2026.

(The $55 billion is the phase-1 budget per SpaceX S-1, with full $119B scale occurring in the mid-2030s, and targeting 1 terawatt of compute per year)

Essentially, Intel would be licensing their fabrication IP and advanced packaging technology for Tesla automotive Chips and Optimus Chips, with the first chips in late 2027 and full-scale production in 2028. Intel signed on April 7, 2026, as the primary foundry partner, contributing the 14A process, advanced packaging, and design/fab expertise (Tesla selected Intel’s 14A for its Terafab chips, and SpaceX is also planning to use it).

This deal is structured so that the $55+ billion used to fund the project comes out of SpaceX/Tesla’s wallet. There is no debt affecting Intel from this agreement. Essentially, Intel is offering their expertise in chip-fabrication/packaging while maintaining 80% margins on their revenue share from Tera-Fab. By 2030, after initial ramping, we model this joint venture to have an ARR of around $4 billion. We believe demand for chips will keep Tera-Fab capacity constrained, allowing Intel to maintain pricing power.

The Economics Behind Tera-Fab:

Total capex spend by 2030 is $55 billion; modelling a fab the size of $55 billion at full ramp can produce up to 52,000 WPM (wafers per month) at 65% ramp capacity; we also project full Phase-1 will be fully ramped and capacity utilized by 2032, with a total projected production capacity of 80,000 WPM, or 960,000 WPY. By 2030 using the base revenue split at 16% allocated to Intel, this leaves them with $4 billion in revenue off of a $25 billion output. There have been specualtions about the revenue split being higher than 13%-16% (around 25%), but this cannot be true as Tesla/SpaceX will operate the FABs themsevles, and Intel would be liscencing their IP. Thus the reason why they can garner 80%-90% margins, but only a maximum of 16% revenue split.

Google TPU Packaging:

The Google TPU deal involves the advanced packaging of 3 million TPU’s for 2028. This will cover roughly half of Google’s TPU supply, with production beginning in 2027. (6 million) As reported, this deal is strictly for advanced packaging, although various sources have stated that it involves fabrication as well. We modelled this strictly for packaging. (If the deal includes chip-fab, the ARR can easily stretch upwards of $5 billion, on a 26% combined margin.)

Using advanced packaging and TPU pricing data, we model this to have an ARR of $2.325 billion, on a 32% operating margin. We landed on these numbers through standard packaging and TPU economics. We predict Intel will charge 70% of TSMC’s CoWoS pricing, as they’ve stated they will offer customers a price-efficiency advantage. (Intel charging roughly $775/unit for 3 million TPU’s equates to $2.325 billion in revenue.) The Google deal was the first large-scale customer Intel landed for its advanced packaging services. This not only benefits Intel by restoring external foundry credibility but also benefits Google by de-risking its reliance on TSMC’s CoWoS and diversifying its supply chain. We thoroughly believe that this will spark more tech giants to adopt Intel’s EMIB technology, given the sheer capacity constraint of the packaging market.

Use this chart below as a reference for TSMC CoWoS bottleneck…

Source: Aria Research

Nvidia Feynman:

Rumours have been circulating that Nvidia is showing interest in Intel’s EMIB technology. The rumours indicate that 75% of Feynman would be allocated to TSMC and their CoWoS technology, whereas the remaining 25% would go to Intel Foundry. Using the 25% number (and also predicting the volume for Feynman), we can allocate a volume of 1.25-1.75 million units.

Using $850/unit and the expected volume, we arrive at a revenue range of $1.04-$1.46 billion. NVIDIA’s higher-end GPU would call for higher pricing than other chips simply because it has a higher bandwidth-bound flagship with next-generation HBM.

This results in Intel using its Foveros 3D die-stacking and co-packaged optics, resulting in a physically larger package than the average chip. Packaging a next-generation chip like Feynman has more bond sites and more bridges and requires 3D hybrid bonding instead of 2.5D. This will lead to more assembly/test work and more yield risk, resulting in a higher service fee for Nvidia.

Apple Fabrication Deal:

The next reported deal is between Intel and Apple. The reporting suggests that Intel will manufacture legacy Apple chips on its 18A process node, with estimates of 15 to 20 million M-series chips a year allocated to Intel, with major ramping beginning in 2027. More recent reporting has Apple showing interest in Intel’s 18A-P node for an entry-level MacBook chip, while also evaluating the 14A process node for more advanced chips. We modelled this to have an ARR of between $1.0-1.2 billion for the entry-level chips. For Apple, this is the beginning of breaking their near “full dependence” on TSMC for fabrication.

This could approximately provide Intel with $1.1 billion on a 22% operating margin. The margin sits below the packaging deal by 10 percentage points, because this is wafer-side work, where costs allocated to process-nodes are much higher. Due to the “crown jewel” of foundry services sitting within chip fabrication, the true value of this deal provides Intel with external fabrication credibility. This validation is even more important, given how supply-constrained TSMC actually is.

For reference, TSMC’s 2nm node is already completely booked through 2028, while Apple currently holds just under 50% of the demand. In layman’s terms, there is no “extra slack” left for Apple to grow their 2nm footprint with TSMC. Currently, TSMC’s 2nm capacity has 78-104-week lead times, pushing new orders into 2028. Even at 100,000 WPM, N2 capacity still remains lower than the combined demand from customers. This can make a strong case for Apple to use Intel’s more advanced nodes in the future for chips such as the A21. CEO Tim Cook specifically stated that they are unable to ramp up supply to meet demand for their advanced chips. Which, in theory, could open up a foundry pipeline between Apple’s advanced chips and IFS. If Intel were to win the A21 chip with average volumes of 40-60 million per year, this alone could supply them with an extra $3.0-$4.9 billion in ARR.

Although this is a possibility, we are purposefully leaving this out of our model as Intel hasn’t fully solidified their advanced nodes as “achieving profitable margins” just yet. Although we believe that the 14A node will be deployed faster than projected (currently 50% yield), the reporting indicates that the A21 will be on TSMC’s N2 node, which makes a lot of sense, given that Apple is a priority customer for TSMC. Overall, at full ramp, the entry-level M-series chips would have an annual revenue run rate of $1.1 billion.

Advanced Packaging Customers:

For Intel, we predict 3 additional advanced packaging customer commitments by 2030, alongside Google and Feynman. We predict the three customers will have an ARR of $2.2 billion. For argument’s sake, we assumed that the customer volume wouldn’t be the same size as Google’s large-scale 3 million orders; we instead modelled it to be around 1.5-2.0 million units per customer.

Large-Scale 14A Customer

In our model, we forecast 3 large-scale customer commitments to Intel’s 14A process node. The revenue per customer sits at $3.6 billion, resulting in a total of $10.8 billion. We modelled this number through thorough estimations of the Wafer ASP for a 1.4nm class-node, along with potential volumes that could be reflective of a “large-scale customer.” Our thesis for Intel’s foundry doesn’t imply the takeover of TSMC on a large scale; it implies companies like Apple utilizing both foundry services as a hedge against a bottleneck within the market. In terms of wafer demand per month, we estimate each customer would command 7,500 wafers per month, roughly, valued at $40,000/wafer. TSMC expects to charge $45,000 per wafer on their A14 1.4nm node. CFO Zisner specifically stated that Intel’s 14A node will have a cost-efficiency advantage while remaining competitively priced. Overall, we believe the $35,000-$45,000 range suffices.

Source: Aria Research

We personally believe that 3 large-scale customer commitments through 2030 are on the more conservative side, and this is done so purposefully to show you the upside potential Intel truly has. Multiple companies have reportedly shown interest in Intel’s 14A process node. Intel has confirmed a partnership with Elon Musk on the Terafab project, while also hinting at several more customers. CEO Tan has also specified that they will not mention the names of any potential customers until they agree to do so. 14A design commitments are expected to start landing in the second half of 2026 and run into the first half of 2027.

The reason these deals have a structural aspect as opposed to a speculative one comes down to the capacity constraint we’ve been preaching about. TSMC’s leading-edge node is effectively sold out, with N2 fully booked. AI buyers are currently trying to outbid one another for allocation. Even Apple, TSMC’s priority customer for over a decade, has dropped below 50% of N2 capacity for the first time as Nvidia and the hyperscalers soak up the node. This forces companies, like Apple, to diversify their supply chain in order to scale supply to meet the red-hot demand. If the demand for XPU’s continue to accelerate (as the data suggests), then large-scale companies that need to increase supply have no other choice but to allocate wafers to Intel’s 14A.

Source: Aria Research

Above are estimations for overall capex allocated to the AI buildout through 2030. Jensen Huang successfully called for $1 trillion in capex by 2026 in the year 2023, and is now predicting 3-4 trillion alone by the end of the decade. Around 40% of all capex is invested into compute, meaning the amount of XPU’s that need to get fabricated will increase alongside Capex.

The Debt Structure:

Inside of the base EPS estimate, we have already included the debt that Intel has incurred PREVIOUSLY in order to buildout their foundry services. The 6% cost of debt comes from the marginal or new-issue rate; in other words, “what Intel would pay to borrow fresh money today at BBB, or a negative outlook.”

Share Dillution:

Before we begin, let’s address the share dillution: Our model assumes that through 2030, Intel will have added roughly 400 million shares to the market to raise capital. In the year 2025, Intel had utilized this method, and the equity structure is as follows…

• US Government Stake: 433.3M x $20.47 = $8.9 billion

• NVIDIA: 214.8M x $23.28 = $5.0 billion

• SoftBank: 87M x $23.00 = $2.0 billion

• TOTAL: 735M Shares for 15.9 Billion Raised

We believe it makes the most sense for Intel to lean heavily on share issuance, as their stock price has received an immense re-rating. In 2025, their share price was hovering around $20- $25 per share, whereas today it’s 5x that level. This means that for less overall dillution they can raise more money, making equity the far cheaper option than debt. If Intel decided to go to the bond market and rely heavily on debt, it would only add costs and harm its balance sheet.

If Intel decides to issue 400 million shares through 2030 (80 million shares per year) at around $100-$110 per share, it could raise $40-$44 billion. This amount of money alone allows them to fully fund the external foundry capacity needed to satisfy our model…

It would cost roughly $20.6 billion to fund the “Aria-foundry model.” Unfortunately, Intel can’t just build a fab that can output 22,500 WPM of leading-edge capacity; Intel must build the whole 40- 50 K WPM fab. (Fab 52 as a reference is 40,000 WPM) The total “funding job” required to build out both fab/packaging facilities at comparable wafer outputs would cost around 30-40 billion dollars. Given this figure, share issuance alone can cover the entire cost. A large caveat that we must cover is the 48D tax credit, from the Big Beautiful Bill…

48D Tax Credit:

The mechanism of the credit allows for a 35% reduction in Capex for qualifying fabs and machinery. The refundable nature of the credit is what makes it very material to Intel. Intel can elect to take the credit as a direct cash payment from Treasury, which means they collect the money even while Intel Foundry is still operating at a loss on paper. A normal tax break is worthless to a company that isn’t paying tax. This one writes a cheque regardless. Essentially, the United States government is co-funding a fab, in cash, with no shares issued, no debt added, and no profit splits. It’s the single most useful funding strategy of their lifetime and puts Intel right in the driver’s seat. The only catch for the credit is that it is only applicable to fabs that break ground before December 31st, 2026, meaning any fab starting construction in 2027 or later cannot reap the tax credit rewards.

In theory, Intel doesn’t even need to raise any debt to fund their foundry, since equity and the 48D Tax Credit are by far the much cheaper options. For our model, we actually assume no new incremental debt will be taken on by Intel through 2030.

Incremental Debt:

So why do we model zero incremental debt through 2030? Because there is nothing left that would require extra debt to fund. The 48D Tax Credit returns 35% of the buildout in cash, cheap equity covers most of the funding required for fabs, and customer pre-payments cover the rest. Debt is the most expensive capital Intel can raise, roughly 6% against an equity cost near 1.2-1.5% and a 48D credit that costs nothing, so a rational management team reaches for it last, and in our model, we believe they never have to.

This doesn’t mean that Intel stops issuing bonds. It will keep rolling over its existing maturities, but refinancing them isn’t incremental debt added to the balance sheet, and the ~$46 billion already pencilled into the base EPS won’t increase any further. If anything, the bigger question our model raises isn’t whether Intel needs debt; it’s whether it even needs all the equity we’ve pencilled in (as cash flow turns positive for the external foundry business, the funding sources start to overlap). Debt only re-enters in our downside case, where weaker cash flows and a pullback in the stock reopen the door to debt over equity. In the base case, it stays at zero, and the balance sheet quietly heals. The bull case is less share dillution, with the implication that the stock continues to surge. (A bull case could warrant 200-300 million shares as opposed to our 400 million model.)

The 1.2-1.5% number represents the earnings yield. , and then using the current Foward PE multiple of 65-85x gives us an earnings yield of 1.2-1.5%. We decided to use the earnings yield over CAPM since CAPM is built specifically for discounting future cash flows and valuing companies. What it doesn’t do well is answer the question of “should I fund this fab with stock issuance or bonds this year.” Comparing that to the established 6%, equity is the much cheaper option to build out their foundry business. Just to reiterate, we used the earnings yield because the equity-or-borrow decision hinges on earnings given up per dollar raised, not on investors required return.

How Much Higher Can Intel Stock Go?

From our homemade valuation model, we have a generous price target set for Intel.

The following images showcase an in-depth analysis of the whole foundry pipeline, along with the incremental EPS each deal provides, EPS before and after financing, and the 2030 “Intel value,” with a 12-month price target.

Read more

Nvidia is shipping. That is no longer the constraint. A finished rack such as the state-of-the-art 800V Vera Rubin or Rubin Ultra rack is just an expensive piece of parts until it is energized, and getting energized now runs through a chain of multi-year queues that the chip cycle has badly outrun. The single most quotable data point in the whole debate captures it: of the roughly 12 GW of 2026 U.S. data center capacity announced across about 140 projects, only ~5 GW is actually under construction, with the rest stuck in the “announced” stage despite normal build times of 12–18 months. Electrical equipment is under 10% of total data center cost and 100% of the bottleneck.

This report is about that bottleneck: what it actually consists of, how long each piece takes, whether it becomes a true ceiling on AI scaling, and which companies are positioned to profit from relieving it. The short answer on winners: the money is moving away from the chip and toward the grid-infrastructure layer. Transmission builders, transformer and switchgear makers, and on-site power generators, because that is where the scarcity now lives.

The Interconnection Bottleneck Deep Dive

The cleanest way to understand why this is so intractable is to see that it isn't one queue: it's four distinct queues stacked on top of each other, each with its own multi-year clock. A project has to clear all four, roughly in sequence, and the binding constraint rotates between them.

Queue 1 — The interconnection study (permission to connect)

This is the grid operator’s process for studying and approving a new connection. In PJM who is the regional operator covering the Mid-Atlantic and the epicenter of the data center surge became so clogged that PJM effectively froze its queue in 2022 to rebuild the entire process, only reopening it in April 2026 with more than 800 projects in progress. PJM has since pushed over 170,000 MW of generation requests through study and is clearing the last ~46 GW of its transition backlog by the end of 2026.

The reformed “Cycle” process promises one-to-two-year study reviews. But two facts gut the apparent improvement:

• A signed interconnection agreement is the midpoint, not the finish. After the agreement comes permitting, financing, equipment procurement, and construction. Realistic commercial-operation dates for projects entering the 2026 cycle stretch into the early 2030s.

• The demand wave keeps outrunning the throughput. PJM expects 30+ GW of new demand between 2024 and 2030, driven largely by data centers. Wood Mackenzie pegs PJM-utility large-load forecasts at ~55 GW by 2030 and ~100 GW by 2037. Some of that is “phantom load” (developers filing the same project across multiple utility queues to hedge their bets) which inflates apparent demand and clogs studies further.

The scale mismatch is structural: utilities designed interconnection processes for 50–100 MW industrial loads. In roughly 15 months, typical hyperscale project size jumped from 100–300 MW to 1 GW or more projects that rival baseload power plants. The planning cycle simply was not built for it.

Queue 2 — Transmission and grid equipment (the physical hardware)

Permission is useless without the tools to move the power. This is where lead times have blown out the most:

• Large power transformers that were able to be obtained 24–30 months before 2020 now stretch toward five years. The big 345kV–765kV transmission-class units run four to five years as well.

• Switchgear is similarly constrained.

• These are globally supply-limited, hard-to-scale factories, and data centers are competing for the same units utilities need for ordinary grid maintenance and replacement.

The transmission build itself including high-voltage lines, substations, and load centers are a specialized construction problem gated by skilled-labor scarcity (linemen and electricians) as much as by materials.

Queue 3 — On-site generation (the main escape, which is also clogging)

Because the grid is slow, developers increasingly bypass it by building dedicated power on-site. But this escape valve is filling up too:

• Gas turbines: GE Vernova’s heavy-duty turbines are effectively sold out through 2030–2031, with lead times around three years. Its combined backlog has blown past 100 GW and is heading toward ~110 GW against only ~10 GW of annual production capacity which is roughly a decade of output with about a fifth tied explicitly to data-center load. Turbine prices are tracking toward roughly triple 2019 levels (~$600/kW by end-2027).

• Reciprocating engines: Caterpillar’s backlog now extends 24 months minimum and is sold out through 2027 even with a major new facility doubling capacity in early 2026; orders placed in early 2026 won’t deliver until 2028, erasing the historical speed advantage of reciprocating engines over turbines. Cummins and INNIO face similar constraints.

The headline number from the field: at PowerGen 2026, one utility planner cited an 84-month (seven-year) lead time to power in a market like Columbus, Ohio, with even the best markets (Pittsburgh, Chicago, Houston, Dallas) requiring 36 months minimum.

Queue 4 — Policy and permitting (the one piece speeding up)

This is the only queue actively improving, under intense political pressure:

• The DOE directed FERC to issue rules accelerating large-load interconnection (FERC historically didn’t even assert jurisdiction over load connections), with action targeted for spring 2026.

• PJM is rolling out an Expedited Interconnection Track for large advanced projects (250 MW+, capped at ~10 per year), a Reliability Resource Initiative (~8 GW across 41 projects), and a “Bring Your Own Generation” fast track.

The flip side is a hardening political consensus that data centers should pay their own way. The White House and all 13 PJM-state governors jointly insisted in early 2026 that data centers bear the cost of their own load growth rather than shifting it onto ordinary ratepayers (echoed by a federal “Ratepayer Protection” framing). And “connect-and-manage” rules increasingly require large loads to accept curtailment (being cut off during grid stress and falling back to on-site backup) as the price of faster connection.

Will this create a point of complete inability to scale?

The scariest outcome would be a point of such saturation that data center growth is forced to grind to a halt as they just simply cannot bring the power to the facility. Limiting growth by so much causing a huge market freefall as most of these companies are propped up on future revenue forecasts.

In good news, all these queues are rate limiting factors, not complete walls to progress. The US has enough aggregate capacity for approved projects it is just the physical hardware (transmission, transformers, switchgear, turbines and the throughput to connect and deliver said power)

None of these are going to be permanently scarce, its just that each of these in itself is a 3-7 year project. So relief along the supply chain is going to arrive slowly but it is expanding as all these manufacturers are pouring billions to maximize capacity as soon as possible.

The realistic outcome is a plateau, not a full stop. Expect a sustained, multi-year period of roughly 2026 through ~2029-2030 where announced AI capacity dramatically exceeds energized capacity, timelines will slip routinely, power is the explicit gating line item in every site decision, and secured firm power becomes a genuine competitive moat. Buildout continues, but at the pace the slowest link allows. This is why when looking at certain company revenue estimates you can see through 2028 most analysts expect lower growth annually as this bottleneck is well understood and accounted for. (For the most part)

Where a true ceiling can occur. Specific regions like PJM's Mid-Atlantic above all, Northern Virginia, parts of Texas can hit effective saturation where no meaningful new large load can be energized for years. That redirects buildout geographically (toward grid headroom, stranded gas, hydro, or wherever firm power exists) rather than stopping it nationally. The constraint reshapes the map of the AI buildout more than it caps the total.

Which Companies are Positioned to Win From Said Bottleneck?

Read more

Micron ($MU) just printed a massive earnings beat across the board, soaring past every single estimate…

• Revenue: $41.46 billion versus $23.86 billion for the prior quarter and $9.30 billion for the same period last year; roughly +346% YoY and +74% sequentially

• Non-GAAP EPS: $25.11; GAAP EPS $24.67 (GAAP net income of $28.24 billion, non-GAAP net income $28.86B)

• vs. consensus of roughly EPS of $20.39 and revenue of $35.1 billion; so a ~$6B revenue beat and a $4.50+ EPS beat

• Board of Directors declared a quarterly dividend of $0.15 per share; ended the quarter with cash, marketable investments, and restricted cash of $30.2 billion

On top of those numbers, Micron increased their margin from 81% - 86% QoQ. This is a company who is printing cash, and those insane margins prove demand is insatiable, and that they have all the pricing power in the world. This earnings call wasn’t just pivotal for Micron as a business, it was a necessity for the AI boom to keep going strong. You see, Micron is one of the global leaders in HBM-manufacturing. HBM is one of the most important bottlenecks in the entire AI stack, and as agentic AI makes its way to the enterprise/consumer level, the need for HBM will only increase. If Micron showed a slowdown in HBM demand, or the beginning telltale signs of an “oversupply,’’ Micron would’ve crashed, and brought the entire market down with it. Micron actually printed the opposite. CEO Sanjay stated that they foresee the HBM bottleneck persisting well past 2027, as well as their pricing power. If you read our previous 2 memos, you would’ve been caught up to speed with this idea. One of our lead AI/Semi analysts, on YouTube and X, has been calling for the HBM bottleneck to persist into 2028 since early April. Once again, this earnings call reiterates this thesis…

Seoul, South Korea:

One of the most random drops we saw in the markets occurred on June 23rd, caused by Korean Debt/leverage. We made a memo regarding the severity of the “leverage issue” sneaking up on Korean investors; primarily retail investors, who have been taking out record amounts of debt in order to invest into tech-heavy companies. Most of these investments being Sk Hynix and Samsung, who make up 42% of the KOSPI index.

Due to the very weak Korean currency (Won), The bank of Korea is planning on raising interest rates toward 3% by September, in order to strengthen it. Seoul, South Korea has already “launched its first joint FX-bank inspection in 14 years, tripled its FX-stabilization bond ceiling, and put the pension fund into $54B of dollar-forward hedging. Tighter Korean policy is contractionary, so the weak won converts straight into the same "rate hike" pressure that's causing de-risking in the leveraged AI trade.”

The MSCI, once again, had excluded Korea from its developed market watchlist, removing the passive-inflow catalyst foreign investors had bought in anticipation of. (About $2.5 billion of foreign capital fled in a single session)

When you have a vast amount of equities being funded by debt, downturns can turn into full on crashes. When and investor borrows money to pay for an investment, the collateral put up is the shares themselves; and when the collateral is losing value, often times, investors cut their losses so they can maintain solvency and pay off their debt. Coupling this with the MSCI exclusion, rate-hikes, and over-leveraging, you have a recipe for a potential market crash. And that’s exactly what happened… for the most part. Leveraged positions like the 2x Sk Hynix and Samsung ETFs both were forced to continuously sell into the close to maintain their multiplier rates. This amplified the sell-off, leading to more and more debt unwinding.

Why this Can Lead to a Relief Rally:

Read more

From claims that Microns earnings will absolutely destroy the entire AI/tech super cycle and lead us into a grave recession to Micron having the best report ever and causing a pump in the market never before seen. Traders are taking very extreme stances on this earnings report.

What we are going to do is cut through the noise and tell you everything you need to know regarding Microns earnings report tomorrow.

Market Expectations

Wall Street is expecting Micron to report revenues of $35.85 billion compared to the $23.86 billion Micron raked in last quarter representing a 50.25% quarterly increase and an EPS estimate of 20.83 vs the reported 12.20 last quarter representing a 70.74% increase from last quarter.

Its clear Wall Street is expecting a huge increase in Microns revenues and EPS and at first glance it can seem like it’s an extreme ask. But once you deep dive into Microns true earning potential it’s clear they will very likely report numbers close to or above these projections.

Now that we know what the market is expecting let’s dive into what our research is telling us and our personal opinions on what the outcome of tomorrows earnings call could be based on our research…

Our Deep Insights into Microns Earnings and Future

We have made it to a point in the AI super cycle where it is clear that memory has become a huge bottleneck. With massive demand and very limited supply that is the core reason behind all of the big memory companies massively increased revenues and EPS.

This bottleneck has given these memory companies all of the negotiating power as they are the ones selling the core piece behind all of the AI hyperscalers ambitions and growth potential. Either they agree to pay whatever memory vendors are charging or they risk falling behind in growth to other companies that will pay that price.

At current estimations we can see this massive pricing power lasting into mid 2027 at the earliest and potentially late 2027 to early 2028 at the latest. However even if supply does increase at that point the prices of DRAM and HBM could and likely will remain elevated from todays prices as demand is unlikely to slow and most likely grow as time goes on.

This will allow all the memory vendors to continue to reach record revenues and earnings for an extended period of time implying a very bullish long term future.

In terms of how much DRAM and HBM are expected to rise; traditional DRAM is estimated to rise ~45% in Q3 with DRAM and HBM combined estimated to rise ~38%. Through Q4 and Q1 2027 DRAM is modelled to rise ~37% total with HBM prices modelled at ~80% increase at the end of 2027 compared to the end of 2026.

All of this is telling you memory is going nowhere soon and will only become a more premium and expensive material in the AI supercycle with the memory vendors continuing to rake in the profits at huge margins.

Earnings Predictions

Read more

If you wanted some red light therapy, you probably decided not to go to your local spa; instead, you looked at CNBC. At the time of writing this article, the Nasdaq 100 is down 2.88%, and the S&P 500 is down 1.30%. Leveraged ETFs like the Direxion 3X Bull Semiconductor ETF are down 21.98%. Overall, not a great day for investing.

You and I alike probably woke up today asking ourselves, “Why are all my positions getting killed?” No need to look any further, because Aria has the answer.

The KOSPI Sell-off

The KOSPI fell 9.99% on June 23rd in an accelerated liquidation. Led by SK Hynix and Samsung (which make up roughly 42% of the index), each dropped around 12% intraday. After a 164% runup during the year, this was the most violent single-session loss of the year. This huge market sell-off was sparked by a few main catalysts that added fuel to the fire.

1. Investors Are Over-Leveraged

   The KOSPI index is the worlds most concentrated expression of the global AI/memory trade, with almost half of the index being SK Hynix and Samsung, the two largest HBM suppliers in the entire world. This in itself isn’t an issue, it only becomes an issue when you take into account how this trade is being funded. Koreans are taking on large amounts of debt to fund their invesments. Korean margin debt hit a record high in June 2026, with specific “leveraged products” being tied directly to the AI/Chip sector. A 3% drop in the overall market could represent a 10% drop in a “debt leveraged” trade; coupling that with high beta AI/Chip companies prone to vast fluctuations, can lead to margin calls at a rapid scale. Margin balances in SK Hynix and Samsung rose from 2.53T won by the end of 2025 to 9.10T won by June 17 2026; roughly 3.6x in 6 months. The collateral put up in these trades are the shares themsevles, so when price falls, the collateral can fall short, leading to margin calls and forced selling of the shares. This already happened recently, with forced sales topping 100B won for five straight sessions (june 5-9), with 169.8B won liquidated on June 9th alone.

2. The Weak Korean Currency

   Coupling over-leveraged margin calls with a depreaciating won, and you have a recipe for an accelerated sell-off. The won currently sits a 1,533 per US dollar; its weakest since the 2009 financial crisis. In order to combat this, the Bank of Korea is planning on raising interest rates toward 3% by September. Seoul has already launched its first joint FX-bank inspection in 14 years, tripled its FX-stabilization bond ceiling, and put the pension fund into $54B of dollar-forward hedging. Tighter Korean policy is contractionary, so the weak won converts straight into the same "rate hike" pressure that's causing de-risking in the leveraged AI trade. When rates rise, the cost of debt does as well, so when this happens, the last thing you need as a leveraged investor is your collateral losing its value. If this does occur, you will be forced by the lendor to liquidate your position.

3. The Japanese Carry Trade

   On top of everything we have discussed, the loaded spring of the Japanese carry trade could release at any given moment. Like Korea, the Bank of Japan is seeing the weakest yen in over 4 decades. Because of this, inflation is running rampid in Japan. This forces the hand of the BOJ to raise interest rates to protect against the inflation issue and strengthen their currency. As we know, the Japanese carry trade has been funding Mega-Cap AI companies in the United States for many years. Insitutions would use this as an “arbitrage technique.” They would borrow yen at an extremely low interest rate, convert it into USD, and purchase treasuries and or stocks that yeild higher returns than the interest they owe.

Looking at the chart above, there is roughly $500b in outstanding yen-funded carry positions worldwide, the majority of which are US equity-related. When rates are rising in Japan, this forces those institutions to adjust their risk. Often times by liquidating the high-beta tech names in order to prevent themselves from being “margin-called.”

What Does This Mean For Your Stocks?

Read more

War Escalation/Recap:

• We saw a slight escalation with the war in Iran, as the Iranian delegation stalled talks with the USA, after renewed threats from the president

• This didn’t sit well with markets, as futures opened down 0.50% across the QQQ and SPY

• After futures opened, reports from various news sources like BRICS News stated that the US and Iran are going to continue talks throughout the week, ultimately entering into a final agreement soon.

• Reports from Iran’s delegation and foreign minister Aragachi claim that Pakistan’s mediation efforts have led to “significant progress toward ending the war in Lebanon,” that war being one of the most material issues for Iran and the progression of peace talks.

What to Expect Today: (with charts)

Read more

So Kevin Warsh just held his first meeting with the FOMC, equity markets sold off heavily into the close interpreting his updated projections as hawkish. However after truly deep diving into the meeting and running our own projections we believe there may be more than meets the eye to the fed meeting and that the market may have grossly overreacted to Kevin Warsh and the FOMCs projections. So lets dive into it to better understand what occurred at the meeting…

The Policy Decision

The FOMC held the federal funds rate target at 3.5-3.75%, the fourth straight meeting with a change. The market did not react well to the updated dot plot (the chart that records where rates are expected to go) these projections turned moderately hawkish with the median year end estimate rising to ~3.8% from 3.4% in March with more than half of the committee predicting a rate hike rather than a rate cut.

At first glance this does sound unfavorable for a rapidly growing AI sector that will rely heavily on low interest rates to speed up growth, but the most important detail was the one that markets overlooked; the long run “terminal rate” which is theoretical equilibrium rate where economic growth is neither stimulated nor inhibited did not change at all; remaining at 3.1%.

Simply put, the committee raised their expectations for the next 18 months but left the endpoint unchanged. The most reasonable read of this info is that the Fed regard the current inflation pressure as a shock we must ride out rather than a permanent shift.

Market vs Fed Expectations

Now diving into how the market vs the Fed is pricing in the future federal funds rate is where it gets really interesting, a popular claim online that the bond market is pricing in three to four hikes is not supported by the rate curve at all. In reality, the market and the fed actually agree on the short term: both of which are implying at least one rate hike toward that 3.8% figure later this year.

However the market and the fed split midway to the end of 2027, where we see the market pricing in additional rate hikes that the feds official projections do not. (see chart below)

So the base case here is that we get one rate hike later this year with divergence through 2027. However this is not set in stone at all as the Fed could potentially reference an alternative inflation measure that could change the projections entirely…

Read more

Its clear everyone’s eyes are on this stock so we’ll break down what is going on in the SpaceX options chain but also examine some of our favorite names in the broader market and we’ll give you exactly what we are eyeing ourselves.

SpaceX -

So first up the big name on the market this week. Options trading just began for SpaceX today and we can already see some pretty incredible things going on so let me break it down for you. First off what is the overall sentiment around this company according to the options market? Short answer: VERY BULLISH!

We saw a call-to-put ratio of 1.4 on this first day which implies a strong bullish sentiment behind the company as the market is very clearly expecting a upward trend in SpaceX. The majority of the volume of calls/puts traded expire July 17 (1 Month till Expiration) with these investors hoping to capitalize on this very volatile stock.

The most popular call strike price is $260 with a volume of 38,677 contracts traded and the most popular put strike price is $135 with a volume of 13,158. So as you can tell the market is expecting HUGE moves in the stock given how wide these strike prices are.

Read more

• We saw huge de-escalation with the war in the Middle East.

• The United States randomly stopped strikes, after it was reported that Iranian Officials offered a framework for a peace-deal, in which the United States had accepted

• The reason it hasn’t been formalized is because of the long transfer times of messages going back and forth, sometimes stretching days.

TODAY:

• After this de-escalation, the need for rate-hikes is almost certainly off the table. Albeit, a formalized agreement materializes

• Watch for a breakout to $725 on the

• If we can sustain $725 in the first 15 we should be good for vast breakout to the upside, post SpaceX IPO

Read more

Nebius -

Nebius was just added to the Nasdaq after hours June 11th, strengthening an already strong stock we’ve been eyeing for months. In particular it will be very important to watch how it moves at open as it is currently up to $236 after hours from its close at $222. Nebius as of June 11th has IV’s for its June 12 contracts at 130-140% so the market is pricing in huge moves of the stock and by June 19th sitting around 120% so if you are a call/put buyer on Nebius you would have to hope it makes a massive move to be profitable.

Overall Rating: Very High IV - High Risk - If you are going to invest in Nebius look to purchase only long dated calls if price is around 230-235 to avoid huge IV crush and theta decay - selling spreads can capitalize on high premiums

Invesco QQQ Trust -

Read more

Introduction

Quantum computing for decades was just a thought experiment; in the 18 months preceding this article, it evolved into something tangible; an engineering program with verified results, peer-reviewed studies, and 10s of billions of public and private capital flowing into this new scientific breakthrough.

However, there is a high amount of confusion regarding what quantum computing actually is and also what industries it can benefit in the future. So in this thesis, we will break down what Quantum Computing actually is, the current trajectory of the sector, and how it all ties in with the current investment landscape surrounding it.

Pillar 1: Science of Quantum Computing

The science of Quantum Computing is extremely complex, and this would turn into a 1000-page textbook if I explained every detail, so I will provide a simplified explanation of the basic principles of Quantum Computing that will become relevant in later parts of the thesis.

Classical Computers store information in bits, which can hold a state of either 0 or 1. Every program ever made runs on this framework. Quantum Computing replaces bits with Qubits, these Qubits can occupy the state of 0 and 1 simultaneously, which are described by two numbers called amplitudes and this state is called superposition. This is the foundation on which all the power of Quantum Computing lies, as it simply allows for an exponentially higher level of computing compared to classical computers for certain specialized tasks.

However, since these qubits exist at the microscopic level, they are incredibly sensitive to the environment around them. Things like temperature changes, vibration, or electromagnetic radiation can scramble their state and result in errors far more often than any typical transistor. The solution to this is Quantum error correction (QEC), combining Logical qubits and Physical qubits to counteract errors from destroying the original encoded data. But how does it do that?

Logical qubits are the software-defined unit of data, and Physical qubits are the physical hardware pieces (like trapped ions or photons) that exist in the processor but are also highly sensitive, as stated. So the single logical qubit encodes its information to many different physical qubits so that if one goes out of whack, the system can recognize and correct the one qubit’s error without having to destroy the original data. A good analogy takes us back to the medieval times, when there was no email, so all messages had to be delivered face to face, but entrusting just one person to travel great distances, kingdom to kingdom, delivering a message was risky because if something happened to them, then the entire message chain collapsed. What they would do is send 10 different people off to deliver the same message to lower the risk of a complete failure. Quantum computing leverages the same principle, encoding the data amongst many different physical qubits so that if one strays, it is easy to correct that error without risking the entire piece of data being lost.

In Quantum computers, each logical qubit requires hundreds to thousands of physical qubits to ensure widespread errors are not a concern. Although this protection mechanism is also the #1 reason behind Quantum Computing’s limitations, the scale needed to minimize errors is huge!

Pillar 2: The Quantum Computing Frontrunners

Now that we have the basic understanding of Quantum computing under our belt, we can get into the fun parts, the companies that are on the front lines of this new emerging technology.

They are separated into 2 distinct categories: The Pure Plays (companies whose entire company is quantum) and the Tech Giants (companies that are cash-rich with great underlying business who are also pursuing Quantum). Let’s start with the Pure Plays.

Pure Players: IonQ, Rigetti, D-Wave, Quantinuum, PsiQuantum

IonQ (NYSE: IONQ) is the largest pure-play by revenue and the first to go public. Its trapped-ion systems reported a record 99.99 percent two-qubit gate fidelity in 2025, and the company has assembled an aggressive roadmap, targeting 10,000 physical qubits by 2027 (following its purchase of Oxford Ionics) and two million physical qubits by 2030. IonQ’s 2025 acquisitions (Oxford Ionics, Lightsynq, Capella Space, Vector Atomic, a majority stake in ID Quantique, and a pending SkyWater foundry deal) extend it into quantum networking, sensing, and security as well as computing.

Rigetti Computing (Nasdaq: RGTI) pursues a vertically integrated, modular superconducting strategy, assembling larger processors from 9-qubit chiplets in its own Fab-1 facility. Its 108-qubit Cepheus-1 system reached general availability in early 2026 at roughly 99 percent median two-qubit fidelity, and its roadmap targets 150-plus qubits in 2026 and 1,000-plus qubits by the end of 2027 at progressively higher fidelity. Rigetti also supports NVIDIA’s NVQLink platform for tying quantum hardware to AI supercomputers, something we will touch on again later.

D-Wave Quantum (NYSE: QBTS) is the commercial veteran of quantum annealing, a specialized approach aimed at optimization rather than universal computing. Its Advantage2 system, delivered through the Leap cloud, underpins real customer deployments, and its 2025 revenue grew 179 percent to roughly $24.6 million at an unusually high gross margin. In 2026 D-Wave also unveiled a gate-model roadmap targeting 100 logical qubits by 2032, signalling a move toward universal computing.

Quantinuum — formed from Honeywell Quantum Solutions and Cambridge Quantum — is the most valuable specialist, reaching roughly a $10 billion valuation in 2025 and filing to go public, with a Nasdaq listing expected in 2026. Its trapped-ion systems are widely regarded as best-in-class on fidelity.

PsiQuantum, still private at a roughly $7 billion valuation after a 2025 round backed by BlackRock, Temasek, Baillie Gifford, and NVIDIA’s venture arm, is making the boldest single bet: skipping the NISQ era to build a fault-tolerant, million-qubit photonic machine using silicon-photonic fabrication with GlobalFoundries, targeted before the end of the decade.

The Tech Giants: IBM, GOOGLE, Microsoft, Amazon

IBM: IBM has published the field’s most detailed and most consistently executed roadmap. At its November 2025 Quantum Developer Conference, it introduced Nighthawk, a 120-qubit processor with 218 next-generation tunable couplers that allows circuits with about 30 percent more complexity than its prior Heron chip and supports up to 5,000 two-qubit gates. IBM projects Nighthawk-class systems reaching 7,500 gates by the end of 2026, 10,000 in 2027, and as many as 15,000 two-qubit gates across 1,000-plus connected qubits by 2028. Alongside it, the experimental Loon processor demonstrated, for the first time in IBM’s line, all the hardware components it considers necessary for fault-tolerant error correction, and the company reported a tenfold speed-up in QEC decoding, a year ahead of schedule, and a shift to 300-millimetre wafer fabrication to double the development pace.

IBM’s stated targets are community-verified quantum advantage by the end of 2026 and fault-tolerant quantum computing by 2029, supported by its Qiskit software stack and a hybrid model that couples classical high-performance computing (CPUs and GPUs) with quantum processors (QPUs). Among incumbents, IBM also has the deepest enterprise-sales motion, which matters more than raw qubit counts for near-term revenue.

Google: Google Quantum AI’s 105-qubit Willow processor produced the field’s defining 2025 result: the first verifiable quantum advantage, using the Quantum Echoes algorithm built on out-of-time-order correlators, published in Nature and estimated at roughly 13,000 times faster than the best classical method for the task. The word verifiable is doing important work: unlike Google’s contested 2019 “quantum supremacy” claim, the 2025 result is reproducible and connects to a real scientific task, measuring molecular structure via nuclear magnetic resonance, which is why it drew broad, if cautious, validation. Google is careful to note this is a specialized tool, not a general-purpose quantum computer.

Microsoft: Microsoft is pursuing the highest-risk, highest-reward path. Its February 2025 Majorana 1 chip is built on topological qubits, which aim to protect quantum information in the hardware itself and, Microsoft argues, could eventually scale toward a million qubits. The science, however, is contested: independent physicists have questioned whether Microsoft’s claims are actually scientifically possible, and rival executives have publicly doubted the result. If Microsoft is right, it could leapfrog the field; if wrong, it has spent years on an unproven foundation. A reader should treat topological qubits as a genuine option, not an established one.

Amazon: Amazon, through AWS, took a more incremental route with its February 2025 Ocelot chip, which combines “cat qubit” technology with conventional error correction in a design that AWS says could substantially reduce the resources needed for error correction and shorten the path to a practical machine by up to five years. Amazon’s commercial center of gravity, though, is its Braket cloud service, which resells access to many hardware partners, a strategy that earns revenue regardless of which modality ultimately wins.

Source: Aria Research

Pillar 3: Use cases, AI + Data Center infrastructure, Quantum Realities

Quantum computing offers an exceptional advantage for certain specialized tasks and no advantage for others. What tasks will Quantum computing actually be useful for? Let’s dive in…

Read more

• Overnight, we saw escalations with the Iran/US war, which led to a final hour of pain in the broader markets

• A United States fighter jet was struck down over the Strait of Hormuz, although Iran denied responsibility

• Trump pledged strikes, causing markets to spook. After trading hours, it was confirmed by CentCom on X that the United States carried out strikes targeting specific locations.

• This sent broader indexes down heavily. The SPY was down an additional $40 after hours, or 0.55%.

• Trump then stated that the strikes were called off after it was reported that Iranian Officials called him pleading to end the strikes. Trump gave them a 1 day extension to sign the deal or be “bombed.”

• This has now lifted markets after hours, leading the QQQ up 0.7%, and the SPY up 0.41%

Jobless Claims & PPI Scheduled for Today:

• The jobless claims are scheduled for 8:30 am today, and the projections are 220,000, with the prior coming in at 225,000

• A blowout beat to the downside will most likely spook markets the same way the labour report did last week. Anything in line or slightly higher will probably help markets recover, giving more fuel to a relief rally on Friday

• In our opinion at Aria, we believe the only catalysts to worry about from the calendar above are the Jobless Claims. Unless PPI is extremely higher than estimates, this shouldn’t move the needle the way the Jobless Claims will.

Key Levels and Stocks to Watch: (AS LONG AS THERE ISN’T AN IRAN ESCALATION)

• Make sure you pay attention to the QQQ’s

• After the dust has settled from the Jobs Report, a recovery of $700 will help push the rally farther. Pay attention to the first 15 mins for the regain of this level. This is crucial for an intraday tech rally.

Read more

We at Aria Research have compiled a group of key catalysts this week. Citing what actually happened, the key hidden details, and what this means for your money. So without further ado, let’s begin…

Nvidia:

NVIDIA kicked off this action-packed week with their keynote speech at Computex 2026. Jensen Huang gave some key insights into the AI trade, boosting specific companies throughout the speech. Take NBIS, for instance. He referred to them as a “world-class AI cloud company,” noting that they are growing extremely fast. Because of this comment, Nebius proceeded to pop 14% on Tuesday, while being up 176% YTD. Marvell also sustained some intense upward pressure after Jensen called them the next “1 trillion dollar company.” Marvell moved 30%+ after the presentation.

Source: CNBC

During the keynote, Jensen kept mentioning that “Compute is revenue.” Now, if you are familiar with our stance here at Aria Research, we are a firm believer in the “capacity constraint moat.” Jensen essentially reiterated what we have been stating for a while. The demand for capacity will be eating up any new energy that comes online for the coming years, with interconnection queues taking 54 months on average. The most qualified and knowledgeable man in the AI space stated that the most scarce product on the market equates to revenue for the masses. Essentially, the companies that can offer quick “time-to-compute” and “compute capacity” will be the winners in this space. If you haven’t yet, please go read the research paper we wrote regarding Iren and Nebius. We explain in great detail the strategic positioning these companies have to gain exponentially from the capacity-constrained environment.

Source: NVIDIA.com

NVIDIA also revealed their RTX Spark Superchip, partnered with Microsoft on ARM architecture. This is a direct shot at AMD, INTC, and Qualcomm, attempting to take market share within the consumer computer sector. One thing Aria flagged that we find interesting is that NVIDIA is dipping their toes in every layer of the AI stack. NVIDIA pushing ARM-based SoCs is a direct shot at x86. Thus, the reason why the three companies mentioned above traded downward after the keynote.

Broadcom:

Broadcom, after hours on Thursday, released their Q2 2026 earnings report.

Source: Broadom Investor Relations

Q2 revenue $22.2B (+48% YoY); AI semi revenue $10.8B (+143% YoY); AI bookings topped $30B. Guided Q3 AI revenue to $16.0B, growth above 200% YoY. The stock printed a monster quarter, but future guidance failed to dazzle Wall Street, even though Q3 revenue guided 29.4B, above 28.5B. Also, fears arose surrounding Marvell and Media-Tek potentially taking market share of their TPU business. Thursday, Macquarie’s downgrade came from their modelling of Broadcom's share of Alphabet's TPU revenue falling from ~95% in 2026 to 80% in 2027 to 65% in 2028, as MediaTek and Marvell gain a role and Google develops in-house capability.

CEO Hock Tan also gave warnings surrounding their margins. Tan's point regarding margins was that segment margins remain very stable and very solid; it's the mix between software and non-AI versus the very rapidly growing AI semiconductor that's diluting gross margin. In other words, the weighted average falls purely because the lowest-margin bucket is becoming the biggest and fastest-growing slice of the pie.

This dilution is mechanical, not a demand signal. Margins aren't compressing because customers are paying less or pulling back; they're compressing because AI demand is so strong it's outgrowing every higher-margin line Broadcom has. If anything, the falling blended margin confirms the boom rather than warning of its end. What actually sold the stock off was a different set of fears: a reaffirmed-but-not-raised full-year AI target into a stock already priced for a raise, and a Google supply chain that's no longer Broadcom's alone. At Aria Research, we’ve modelled demand to keep booming until at least 2029 as a base case, with the bull cases exceeding 2030 onwards. The data backs up demand only accelerating, not softening, which is why we called this a textbook “buy the dip” opportunity on Thursday.

Intel:

Intel kicked off their action-packed week with their Keynote speech at Computex 2026. Firstly, Intel confirmed their next-gen Xeon 7 chip, “Diamond Rapids,” to be manufactured on their 18A-P node in 2027, which is extremely positive. Many analysts have written about Intel having their yields at 50%, raising heavy issues regarding the future of its foundry. CEO Tan has consistently reiterated that they are increasing yields at 6-8% per month, and their Diamond Rapids chip got rid of any worries regarding yields and future node capabilities, in our opinion.

Source: Intel.com

Intel also leaned heavily on the “AI-everywhere story,” reiterating x86’s deep legacy as the architecture leader, as well as the CPU revival story. Intel presented a live demo of inference requests on a program, showing that the GPU/CPU ratio could reach 1:1, as agentic AI scales and becomes more prominent. Intel had left out any key information regarding their foundry, which is what investors care about most. At Aria Research, we believe more deals are inevitable for Intel’s foundry, as it is extremely bottlenecked. Due to the lack of “foundry talk,” Intel stock saw a pullback.

The Great Market Crash of June 5th 2026:

If you wanted to do some red light therapy, looking at the stock market on Friday last week was the best way to do it. The QQQ was down -4%, the S&P 500 dropped around -2.6%, resulting in 1.5 trillion dollars being wiped from the stock market. This was the first losing week in 10, and the biggest drop since October 10th. The trigger? The labour market, and SpaceX ipo.

Employers added 172,000 jobs in May, more than double the estimate of 80,000. The reason this killed markets is that it got rid of any potential for a rate cut. When the labour market is running really hot, it leads to more money circulating in the economy, resulting in more consumer spending (demand), increasing prices… or, in layman’s terms, inflation. Investors hate inflation, and the only way to combat it is to raise rates. The expectations for a rate hike rose to 66% on Polymarket. This is so pivotal for the AI trade because companies are borrowing a large amount of money for expansion, and when rates increase, it artificially cools growth and the economy. This spooked investors and made them sell out of stocks at an alarming rate. Great companies that we’ve flagged dropped 10%+, such as AMD, INTC, NBIS, IREN, and MRVL.

An important part of this sell-off was the SpaceX ipo. We saw a lot of money flow out of the stock market and into T-bills and sidelined cash. This is extremely important as no money flowed into precious metals like gold and silver. When money flows into gold and silver, this means institutions believe that a recession or large-scale crash is coming. Since this didn’t happen, we can rest assured this was more of a “one-off-flash crash.” The largest IPO ever is occurring on Friday, June 12th, offering at $135 per share, and a 1.77 trillion market cap.

A “once-in-a-decade” offering can make any investor pull sidelined cash and invest in it, and SpaceX is exactly that. When money moves into cash/cash equivalents, it means people are most likely preparing to unleash it. And the timing makes perfect sense. The “dry powder” is unambiguously there, the largest ever, and the SpaceX raise is trivially small against it, so a relief rally is well within the realm of what this cash could fund. What the data does not show is that the cash is already turning over. What this means is that a solid SpaceX IPO could be the trigger for a relief rally in hardware stocks in particular.

Below is a chart showing the amount of sidelined cash and comparing SpaceX's funding against the total cash value. (Split roughly 3 trillion in retail, and the rest institutional)

How to Strategically Play the Upcoming Weeks:

Encompassing everything that occurred, we have learned a few things that need to be addressed, so that we can make informed investment decisions…

Read more

A few years ago, the AI trade was dominated by compute. Companies couldn’t get enough of it. The clear winner? Nvidia. They made the best product tailored to training LLM’s like ChatGPT and Claude. From there, the shift went from training AI to implementing it at the enterprise level. This shift is called Agentic AI, and this is the future. It allows businesses to have artificial agents work alongside employees to streamline operations. Whether it be scanning basic IT tips, reading charts, and analyzing data, these agents can perform tasks effectively.

The caveat is that agentic AI is creating a vast amount of demand for the entire AI stack. That being semi conductors, high-bandwidth memory, NAND storage, advanced packaging, chip fabrication, cloud services, and data centres. The issue is that demand is so strong for the stack that supply simply cannot keep up. The CPUs, GPUs, storage, and memory are being placed into data centres in order to store, process, train, and implement advanced AI models and agents. This is where the bottleneck lies.

The demand for data centres is so vast, and the main players in this trade are all expanding capacity. Albeit, it takes 2+ years from ground being broken to having a fully functioning data centre. These data centres can consume as much power as a small city. Lead times to get energy are sitting at 4-5 years, as grid interconnections are severely backlogged. Everyone is desperately trying to get power, but they simply can’t. There is roughly 2,000+ gigawatts of backlogged energy from interconnections that won’t come online until 2030-2033 (as of May 2026). There is more energy backlogged than the entire United States Grid has online currently.

This leaves hyperscalers and AI companies with two options.

1. They can wait patiently to have more capacity come online, and risk missing out on revenues and potentially falling behind in the “AI race.”

2. They can purchase capacity from third-party vendors who offer it.

This is where the thesis of our memo comes in. Neoclouds like Iren and Nebius are well-positioned to benefit from capacity scarcity and the overall energy bottleneck. Iren, formerly known as Iris, used to be a bitcoin mining company. They own the full stack, meaning they own the land, the data centre, and the compute within it. Iren has 4.5+ GW of secured capacity, which can power roughly 4.5 million average homes. Due to their past, they locked in grid capacity at low premiums and are able to rent it out to companies at a larger premium.

Nebius, on the other hand, was formally the international division and Dutch parent company of Yandex, known as the “Google of Russia.” After the Ukraine war began, they sold all of their Russian assets and pivoted to being a full-scale artificial intelligence company. Similar to Iren, they currently rent their capacity. These two companies alike can benefit immensely from the energy constraint that will be plaguing hyperscalers and blue-chip companies begging to win the AI race.

The Following displays Nebius and Iren’s Market Price at the time of this Memo…

Pillar 1: The Grid Constraint Explained

The expansion of data centres and the need for agentic workloads will only constrain the grid further. Every data centre turned online requires roughly 20-100 megawatts of power (0.02-0.1 gigawatts). This is enough to power 10,000-100,000 households. One data centre being deployed uses as much energy as a small city. This is clearly a problem, especially when the United States grid can support only 1,250 Gigawatts of energy as of today. As mentioned above, around 2,000 gigawatts worth of backlogged energy will be deployed fully between 2030-2033, as a mix of solar power and renewable sources, nuclear, and gas turbines, in order to support the expansion of artificial intelligence.

Lead times stretch from 4-5 years, as the queue to acquire power is extensively backlogged. The simple answer. Interconnections.

Read more

Intel is no longer the same company it was eighteen months ago, and the market is now beginning to realize this. With an extensive history of failed execution and collapsed foundry practices behind them, they have successfully put together a catalogue for the first time that merits a reevaluation. The new CEO, Lip-Bu Tan, is on board with changing their companies structure. Alongside the US government and Nvidia, which have taken an equity stake in Intel. As such, their 18A foundry technology is now commercially shipping products with yield improvements being greater than what their internal projections were at. Advanced packaging is also growing into a $1B+ business, and companies that use external foundries are making significant strides in committing to Intel long term.

This stock has already had its fair share of upside (approximately 400% over the last twelve months) as markets begin to digest the operational data that we have been seeing. The question at hand is whether Intel has already “topped out” or if there is any more upside left in the stock. In terms of a baseball game, I believe we are in the early to middle innings in the Intel bull cycle, with the majority of the remaining catalysts yet to come.

The thesis rests on six pillars: foundry customer commitment, their advanced HBM packaging (EMIB and EMIB-T), the geopolitical “floor” as defined by the U.S. Government’s stake in Intel, the roadmap of 18A and 14A technologies, the Lip-Bu Tan operational reset, and the capital expenditures peak that will lead to an inflection point in free cash flow in 2027 & 2028. Altogether, I believe these pillars create immense upside potential for Intel that some analysts are failing to capture.

The Setup: Where Intel Stands Today

Intel’s Q1 2026 revenue was $13.6 billion, up 7% year over year. The Data Center and artificial intelligence revenue increased by 22% to $5.1 billion in Q1 2026. The revenue from Intel Foundry services increased by 16% to $5.4 billion in Q1 2026. The revenue from Client Computing remained steady at $7.7 billion in Q1 2026. On April 18, 2026, Intel issued guidance of between $13.8 billion and $14.8 billion for its second quarter (Q2) earnings. The stock moved 24% in one day following Intel’s reporting of its first-quarter results.

While the headline numbers were important, they were much less important than the shift from the bear case on Intel, which is now required to address specific execution failures rather than vague concerns about strategic direction. Before this time, the bear case was simply “that Intel was a melting ice cube.” Today, the bear case is simply that the recovery has already been priced in. As such, these are very different problems for investors to underwrite, indicating a strong narrative shift on Intel.

Three structural changes happened between late 2025 and early 2026 that the market is still digesting:

1. The US government took a 9.9% equity stake in August 2025 at $20.47 per share, funded through CHIPS Act grants. The total government investment now stands at $11.1 billion, with a five-year warrant for an additional 5% stake exercisable if Intel ceases to own at least 51% of the foundry business.

2. NVIDIA invested $5 billion in September 2025 at $23.28 per share, completing the transaction in December after FTC clearance. The deal includes a joint product roadmap covering custom data centre CPUs and x86 SOCs with integrated RTX chiplets using NVLink.

3. Apple has reportedly signed an NDA with Intel and received the 18A-P PDK 0.9.1GA, with Ming-Chi Kuo estimating production silicon shipping in Q2 or Q3 2027 at volumes of 15 to 20 million entry-level M-series chips annually. The Wall Street Journal confirmed a preliminary agreement in early May 2026.

These three developments anchor the bull case in a way that pure operational improvement cannot. The political, ecosystem, and customer signals are now aligned in a way they have not been since the early 2010s.

Pillar One: The Foundry Customer Commitment Ramp

This is one of the least discussed legs of the thesis since the financial impact will take time to come from this announcement cycle. The external foundry revenue in Q1 2026 was only $174 million. As such, that is a very small percentage of the total foundry revenue at $5.4 billion. Meaning, the majority of foundry revenue comes from internal chip production.

The number bulls should be watching is different. On the Q1 2026 earnings call, Tan disclosed that Intel signed multiple long-term agreements in the quarter for volume and pricing of server CPUs and ASICs, typically lasting three to five years, with several confidential at customer request. The publicly named deal was a multi-year LTA with Google covering both Xeon CPUs and ASICs. NVIDIA also selected Xeon 6 as the host CPU for its DGX Rubin NVL8 systems. The previously confirmed Microsoft custom 18A chip deal remains in development, with Amazon’s AI fabric chip and the Terafab deal with Tesla, SpaceX, and xAI rounding out the publicly disclosed customer set.

CEO Tan has stated that design agreements from external foundry customers will start in the second half of 2026 and will expand throughout 2027. As well, CFO David Zinsner reiterated this, stating that “signals” from customers (outside) will “become more concrete” during the second half of 2026. While any individual name is less important than the overall trend, multiple foundry customer wins accumulate. Winning the first major customer is difficult because you are asking an external customer to invest in a multi-year product development cycle based upon new, unproven manufacturing technology. The Apple deal, if it converts from a preliminary agreement to a signed wafer agreement, would be the validation that unlocks the next wave of customers waiting on the sidelines.

The financial inflection will occur when the company’s external foundry revenues scale from under $200 million on a quarterly basis to $1 billion or more per quarter. This would cause the foundry portion of the business to go from being a $2.4 billion per quarter money pit with no operating income to approximately break-even. In theory, this could happen as soon as the fourth quarter of 2027 or the first quarter of 2028; however, investors usually price it in around twelve to eighteen months before it actually happens.

Read more