One thing worth saying plainly up front: I’m not a financial advisor, and neither is my partner. Everything here comes from personal interest, not professional training — I just happen to enjoy thinking hard about how to use the different account types available to my household as efficiently as possible, especially when it comes to taxes. Treat this as one household sharing its reasoning, not a recommendation.

The 401k: simple, on purpose

The verdict: broad, aggressive, and boring by design.

One account, split 60% S&P 500 / 40% Nasdaq 100. I’m still young enough to ride out volatility, so I lean aggressive now and plan to dial it back as I get closer to retirement. The reasoning is simple: the US stock market has been the engine of long-term growth, and tech has been the engine within that engine. I’m not trying to be clever here — just consistently exposed to it.

Why I prefer Roth over Traditional — for (almost) everyone

The verdict: Roth usually wins once you actually compare tax brackets now vs. later.

The Roth-vs-Traditional debate comes down to one question: is your tax bracket higher today, or will it be higher when you withdraw? For me, the answer points to Roth, for a few reasons:

• I don’t plan to spend luxuriously now and cut back later — I’m aiming for a steady lifestyle whether my income goes up or down. That means my spending, and the taxes tied to funding it, don’t meaningfully shrink in retirement.

• Inflation alone pushes future expenses up, and tax brackets have historically trended up over time, not down, for the same real income.

• If I ever end up leaving the US — whatever the reason — withdrawing a Traditional account as a lump sum triggers a steep tax hit — a large withdrawal in one year can jump you into a much higher bracket. A Roth withdrawal, lump sum or not, is tax-free. That flexibility matters a lot for a family that isn’t guaranteed to stay in one country long-term.

  Subscribe now

Why I invest aggressively inside the Roth specifically

The verdict: tax-free growth and high-risk, high-reward picks are a natural pair.

If withdrawals are never taxed, the incentive is to make the growth inside the account as large as possible. That’s why my Roth IRAs are managed more aggressively than my 401k — the benchmark I’m trying to beat isn’t the S&P 500, it’s the Nasdaq 100. That’s a deliberately high bar, and I track myself against it directly.

How I pick stocks

The verdict: growth with real fundamentals behind it, bought at a price I think undersells the business.

I’m not indexing inside these accounts — I’m picking individual names. The filter I use for every position:

• Growth-stage companies with real fundamentals behind the growth

• Mispriced relative to what I think the business is worth

• Strong management teams with a track record

• Healthy balance sheets — I’m wary of growth funded by heavy debt

• Decent momentum, but not stocks that already look overextended

Two accounts, two risk levels

The verdict: same target, different risk budget, because one account has backup and the other doesn’t.

Between the two Roth IRAs in my household, I’ve started calling them The Comet and The Oak — those names capture the actual difference better than “account one” and “account two” ever did. The Comet is the one that also has a 401k sitting behind it, so it can afford to run hot — there’s other tax-advantaged savings backing it up if it has a rough stretch. The Oak is the one that’s the sole retirement account for whoever holds it, so it’s built to be steadier by design, not by preference. Both are still measured against the same target: beat the Nasdaq 100.

One structural note worth mentioning: Roth IRA contributions are capped by the IRS — $7,500 per person, per year, for 2026 (under age 50). That’s a real constraint on how fast these accounts can grow from new money alone; the rest of the growth has to come from what I actually do with what’s already inside.

One more thing worth flagging, especially if your household income is on the higher side: Roth IRA contributions also phase out above a certain income level — for 2026, that’s a modified adjusted gross income (MAGI) of $153,000 for single filers and $242,000 for married couples filing jointly, phasing out completely a bit above those numbers. If you’re over the limit, you’re not locked out — a backdoor Roth IRA (contributing to a Traditional IRA, then converting it to Roth) is a common, IRS-sanctioned workaround. It’s a bit more involved procedurally, and worth understanding the pro-rata rule if you already have other Traditional IRA balances, but it’s the reason “I make too much for a Roth IRA” usually isn’t actually true anymore.

What’s next

The next two articles go through each account in full: every position, why it’s there, and how each account is tracking against the S&P 500 and Nasdaq 100. Not a stock tip sheet — a running, honest record of what I did and whether it actually worked.

Subscribe now

Disclaimer

This article is for general informational and educational purposes only and does not constitute financial, investment, tax, or legal advice, nor a recommendation or solicitation to buy, sell, or hold any specific security mentioned. The author is not a licensed financial advisor, broker-dealer, or registered investment advisor. All performance figures shown are historical, self-reported, unaudited, and not independently verified; past performance is not indicative of future results, and no representation is made that any reader will achieve similar returns. Individual investments and portfolios discussed reflect the author’s personal risk tolerance, timeline, and circumstances, which may be entirely inappropriate for other readers. All investing involves risk, including the potential loss of principal. Index comparisons (S&P 500, Nasdaq 100) are for illustrative context only and do not account for differences in risk, fees, taxes, or investment objectives between the author’s holdings and those indices. Readers should conduct their own due diligence and consult a licensed financial advisor, tax professional, or attorney before making any investment decision. The author and this publication expressly disclaim any liability for losses or damages arising from reliance on this content.

I mapped it out as twelve stages, because that’s genuinely how many distinct things have to happen, in roughly this order, before any of this spending turns into a working AI system.

The AI buildout isn’t one trade. It’s an investable supply chain twelve layers deep — and the layer that gets covered to death (chips) is maybe the third-most interesting one in it.

So here’s the question I actually want to answer in this piece: which stage is holding up the rest right now?

Not chips. That was the answer through roughly 2023–2024, and it’s when most of the current investing consensus about this whole space actually formed. Today the tightest constraint is power — specifically, getting a data center connected to the grid, and the electrical equipment that connects it. That’s stages 02 and 04 below, and I don’t think most portfolios are positioned there at all.

But I think there’s a more useful distinction than just naming which stage is stuck:

Some constraints, capital can eventually clear. Some, it genuinely can’t. The market tends to price them the same way, and I think that’s the mistake.

Here’s what I mean. A memory shortage is fundamentally a capital problem. High-bandwidth memory is effectively sold out and advanced packaging is booked solid — but Micron, SK Hynix, TSMC, and Amkor can all respond by building more of it, and they are, at enormous scale. It’s brutally expensive and it is not fast — more on the actual timeline below — but there is a path where money solves it.

An interconnection queue is not a capital problem. Neither is a permitting office, an environmental review, or a transformer factory whose output is limited by how many transformer factories exist in the world. You can’t write a check to move faster through a queue that’s regulatory in nature and staffed by people who don’t work for you.

That distinction — clearable with money versus not — is probably the single most useful lens in this whole series, and it’s why I think the “boring” stages end up carrying the real pricing power. Worth keeping in mind through everything below.

Each stage gets a rough read on where it stands now, and — where I could actually pin one down — when the constraint eases.

Subscribe now

01 — Raw materials. Constrained, and I don’t see a clear resolution date. Before a single chip exists, someone has to dig up the rare earths used in magnets and motors, refine the silicon, and secure specialty materials most investors have probably never heard of — helium, ABF substrate film, T-glass fiberglass — that have quietly become real 2026 bottlenecks. The 2021–22 chip shortage was fundamentally an inventory problem, and inventory problems resolve once the overordering works its way through the system. This one isn’t like that. These materials get consumed about as fast as they’re produced, with no warehouse cushion sitting around to bail anyone out. Partially clearable, but on a mining timeline — which is to say, slowly.

02 — Energy and grid. Active bottleneck, and I’d call this the binding constraint of the whole cycle right now. A data center is really a power problem wearing a computer’s clothing. Gas turbines, nuclear, and solar-plus-storage are all competing for the same limited pool of construction capacity, because these are the longest-lead assets anywhere in the chain.

Then, once the power exists, it actually has to reach the site — and that’s where the buildout hits its hardest wall. The interconnection queue (the line utilities keep of projects waiting to be formally connected to the grid) is genuinely staggering right now. Lawrence Berkeley National Laboratory’s most recent review counts more than 2,000 gigawatts of generation and storage waiting for grid access — more capacity than the entire United States currently has installed. The typical project now waits about five years from application to commercial operation, up from roughly two years two decades ago. Historically, only about 19% of projects that entered these queues between 2000 and 2018 ever actually reached operation.

In fairness, there’s one genuine bright spot: Berkeley Lab found total queue capacity declined in 2024 for the first time, as reforms began clearing out speculative applications. But the direction of travel for wait times hasn’t reversed. Largely unclearable by capital. Very investable as a constraint, if you know where to look for exposure to it.

03 — Site and shell. Ongoing, project by project — permitting is really the swing variable. Land gets bought, and then bureaucracy starts. Permitting and environmental review can eat up as much calendar time as actually constructing the building. This is the stage where “the capital is committed” and “the facility is open and billing customers” quietly stop being the same sentence.

04 — Facility fit-out. Active bottleneck, and I think the tightest calendar mismatch in the whole chain. The empty shell gets wired, cooled, and connected — and this is where the numbers get genuinely absurd.

Wood Mackenzie’s Q2 2025 industry survey — the most recent detailed order-based data I could find — put average lead times at 128 weeks for a large power transformer (about two and a half years) and 144 weeks for generator step-up transformers, the units that connect a power source like a turbine into the grid. Some specialized orders stretch beyond four years. Prices are up 77% since 2019, with demand for high-voltage transformers up 116% and for generator step-up units up 274% over the same period.

Two details that make this worse than the headline numbers suggest. Roughly 80% of large power transformers used in the U.S. are imported. And the specialized steel they require — grain-oriented electrical steel — has exactly one domestic producer, Cleveland-Cliffs.

I’ll flag the one piece of good news honestly: smaller distribution transformers have improved substantially, down to a fraction of their 2023 peak wait times. It’s the large units, the ones a data center actually needs, that remain stuck.

A typical AI deployment is supposed to happen in under 18 months. That’s a multi-year part trying to feed an 18-month plan. Wood Mackenzie’s own framing is the part I keep coming back to: for large industrial developers, equipment availability has now replaced capital and permitting as the primary constraint on project timelines. That’s about as direct a statement of this entire series’ thesis as I could ask for, and it didn’t come from me.

05 — Chip design and fab equipment. Mature, well covered, and mostly already re-rated by the market. Electronic design automation (the software used to actually design a chip before it’s manufactured) has become one of the more defensible business models in all of technology, and the equipment makers who let a foundry manufacture what’s been designed are just as entrenched. Both are excellent businesses — which is exactly why this corner of the map is among the more efficiently priced.

06 — Memory and fabrication. Constrained, sold out, and clearable in principle — but not before 2028. This is the stage where I had to correct my own thinking while researching it.

The shortage is real and well documented. SK Hynix has confirmed its entire 2026 HBM capacity is booked, with the crunch extending into 2027. Micron has said the same. TSMC’s advanced packaging capacity — the process of sealing a finished chip into something usable, where the memory physically gets mounted alongside the processor — is fully booked as well. A manufactured chip with no packaging slot is basically a finished car with nowhere to park it.

Here’s the part I initially got too optimistic about. Capital is responding, at genuinely enormous scale — Micron, SK Hynix, and Samsung are all building. But Micron’s new fab doesn’t produce until roughly Q3 2028. SK Hynix’s capacity additions are reportedly committed through 2028 as well. Industry estimates still show a meaningful memory shortfall at the end of 2027. So “capital can clear this” is true, and it’s a real difference from the power constraint — but the relief arrives in 2028, not next year.

One detail I find telling: hyperscalers have reportedly put billions in prepayments down to lock in future memory supply. That’s the same behavior Part 0A found in the lease footnotes — commitments being made years ahead of anything showing up as spending.

This is also where the capex dollars concentrate hardest: $380 billion of the $725 billion, per Part 0A’s allocation. Note the asymmetry — the tightest constraint in the pipeline is also the most fundable one. Amkor’s $7 billion Arizona campus is exactly that response. It produces nothing until 2028.

07 — Servers, optics, and networking. Scaling fast, only moderately discovered by investors so far. Individual chips get bolted into servers and racks, then wired together with optical interconnects and rack-scale networking so that thousands of chips can behave like one machine.

08 — Neocloud and compute operators. A young category, and financing terms are actively repricing right now. These are companies that buy or lease enormous GPU fleets and rent the compute back out — a layer most retail investors have genuinely never priced, because it’s newer than the rest of this list. It carries different financing and different credit risk than a hyperscaler running its own infrastructure, and lenders are now pricing that difference explicitly. Recent data center debt has cleared at spreads meaningfully above comparably rated debt with no AI exposure attached. That spread is basically the market’s live opinion on how risky this stage actually is.

09 — Cluster software. A mature enterprise software cycle, tracking the hardware buildout with a lag. Firmware, orchestration, and distributed training frameworks — the unglamorous layer that makes tens of thousands of chips actually behave like a single machine instead of a pile of expensive parts.

10 — Data pipeline. Ongoing, and there’s genuinely thin public-market exposure here. Someone licenses and collects the raw material a model learns from, then cleans, deduplicates, tokenizes, and stores it at a scale that’s becoming a real driver of enterprise storage demand in its own right.

11 — Model development. Ongoing, and almost entirely private or in-house. Pretraining, fine-tuning, RLHF (reinforcement learning from human feedback, one of the main techniques used to shape a model’s behavior), evaluation — done almost exclusively inside hyperscalers and frontier labs. Not a category with many investable pure plays, and I don’t think that’s likely to change soon.

12A / 12B — Deployment, splitting in two directions. Cloud deployment is monetizing right now; physical deployment is earlier and more speculative. A finished model either goes out through the cloud, where inference serving and enterprise platforms turn it into something a business actually pays for — or it goes into the physical world, where edge chips, robotics, and autonomous systems put it to work directly.

Two things that don’t fit cleanly into a single slot, worth flagging on their own. Cybersecurity isn’t really a stage — it’s more like a live wire running underneath the neocloud layer, cluster software, and cloud deployment all at once, the moment any of that infrastructure actually goes live. It gets its own treatment in Part 3. Quantum computing sits next door to this whole pipeline rather than inside it: no real dependency on the rest of the buildout, no shared revenue driver, and a meaningfully earlier adoption curve. Anywhere it comes up later in this series, I’ve flagged it as speculative — not as a hedge, just because that’s the accurate description of where it actually is.

The scoreboard

Twelve stages, twelve different clocks.

Five of them are genuinely stuck. Two are constrained with no clear resolution date I could find (raw materials, memory). Three are active bottlenecks with a dated but distant end point (energy and grid, site and shell, facility fit-out). Two are mature and already efficiently priced (chip design, cluster software) — which is exactly why neither shows up as a fresh opportunity anywhere in this series. The rest are scaling, repricing, or genuinely mixed.

Now apply the lens from earlier. Of those five stuck stages, only one — memory — is a constraint that capital can actually clear, and even that one doesn’t clear until 2028. The other four are limited by mining timelines, five-year interconnection queues, permitting offices, and transformer factories. Money doesn’t move those. Time does.

I think that spread is the whole reason a single “AI stocks” bucket has never really made sense to me as a way to think about this.

Subscribe now

Why the order matters more than the labels

A bottleneck at an early stage strangles everything built on top of it, no matter how much capital or demand is stacked further down the chain. You can own the best inference platform in the world and still be sitting there waiting on a substation.

The stages that sound the most boring — raw materials, energy and grid, site and shell — are, I think, frequently the ones actually setting the tempo, and the pricing power, for everything downstream that gets the headlines.

And the bottleneck moves over time. It was chips. It’s power now. When the power constraint eventually eases — probably late this decade, if the queues clear on anything like their current pace — I’d expect it to move again, likely to whatever’s scarcest at the moment everything else has become abundant. I think tracking that migration is a more useful exercise than trying to pick a single winner inside any one stage.

It’s also worth being straightforward about where public markets simply don’t reach. Data collection, pretraining, fine-tuning, evaluation — all real, all necessary, all dominated by private vendors and in-house lab teams rather than anything you can actually buy a ticker for. Naming that gap honestly feels like part of giving you the real picture, not a footnote to skip past.

The mistake I made in my own research, and want to flag

Starting with Part 1, every subsector in this series gets two separate reads, kept deliberately apart.

First: what phase of the physical buildout is this actually in. Second: what phase is the stock trading at.

Those two questions move independently, and I think blurring them together is the single biggest error I made earlier in putting this project together. A real, bottlenecked, genuinely early-stage sector can be attached to a stock that’s already up 500% and priced for perfection — and the reverse happens just as often, in the other direction. Keeping “where’s the buildout” and “where’s the stock” as two separate questions is really the whole point of doing this properly, and it’s the lens every table from here forward is going to use.

The map for the next four parts

Stages 10 and 11 — the data pipeline and model development — don’t get their own part. They have essentially no meaningful public-market exposure, so I’ve folded them into the context of other parts where they’re relevant, rather than pretending there’s a standalone investing angle that doesn’t really exist.

Three things you won’t find in this series, on purpose

Mega-caps. Nvidia, TSMC, Broadcom, and AMD touch nearly every stage in this chain, and they’re also probably the four most-covered stocks on the planet. Leaving them out isn’t a judgment on how important they are — it’s just where I think this series can actually say something the market isn’t already pricing efficiently.

Predictions. Every part in this series describes where a stage, and its related stocks, stand today. Never where I think a stock is going next. That’s a line I’m being deliberate about not crossing.

Momentum names riding pure sentiment. There’s a separate, clearly-labeled section later in this series covering the social-media-driven names closer to publication — framed honestly as sentiment plays, not fundamentals plays, because that’s what they actually are.

Twelve stages. Five stuck. One of those five fixable with money, and not until 2028.

The other four are where I think this series is going to spend most of its time, because that’s where a constraint genuinely can’t be competed away with capital — and where I think the market is still pricing this like a chip cycle, when it’s actually behaved like a construction cycle for a while now.

Let’s go stage by stage.

Disclaimer

This article reflects my own research into public company filings and industry data, and is provided for general informational and educational purposes only. It is not financial advice, and I am not a financial advisor. Nothing here is a recommendation to buy, sell, or hold any security mentioned. Key figures are sourced from Wood Mackenzie’s Q2 2025 transformer survey and subsequent commentary, Lawrence Berkeley National Laboratory’s “Queued Up” interconnection review, public statements and earnings materials from SK Hynix, Micron and TSMC on memory and advanced packaging capacity, Amkor Technology disclosures, and company Q2 2026 earnings materials. Some industry allocation and capacity figures circulating publicly trace back to a small number of sell-side estimates rather than independent measurement, and I’ve tried to keep those qualitative rather than precise where that’s the case. The status label attached to each stage is my own categorical judgment based on available reporting, not a precise or official measure, and conditions in any single stage can shift quickly — these are current only as of the date noted at the top. This is a passion project sharing my own research process, not investment advice tailored to your situation — consult a licensed financial advisor before making any investment decision.

I’ve spent the last few weeks going through the actual filings behind the AI buildout — not the headlines, the footnotes. This piece is the result, and it’s the reason I split what was going to be one overview into two: this one follows the money itself, and Part 0B (coming next) walks the physical stages that money has to survive.

On the Fourth of July, in the Idaho desert, a startup called Aalo Atomics brought a nuclear reactor to criticality against a presidential deadline — one of four to do it that month. It doesn’t generate electricity yet. What it generates is a date: a plan to have a data center running directly on top of it, on site, next year. Not connected to the grid. Sitting on the reactor, because the grid can’t reach these buildings fast enough to matter.

Seven hundred miles due south, in Peoria, Arizona, a $7 billion packaging plant is going up — not because anyone doubts chip demand, but because the machines that seal finished silicon into usable form are the scarcest link in the chain. It doesn’t start production until 2028.

Everyone’s debating whether Nvidia is overvalued. I think the more interesting argument is sitting in lease footnotes, equipment backlogs, and utility filings that most investors never actually open.

So let’s follow the paper.

The verdict: the number every headline quotes is the floor of the real commitment, not the measure of it — and I think that’s the single most useful thing to carry into the rest of this series.

Four companies, three-quarters of a trillion dollars, one year

Amazon, Microsoft, Alphabet, and Meta guided to a combined $725 billion of capital expenditure (”capex” — money spent building or buying long-term physical assets like data centers and chips, as opposed to day-to-day operating costs) in 2026 — up 77% in a single year, and roughly four times what the same group spent in 2022.

Then, in nine days at the end of July, all four reported earnings. Amazon raised its number to $220 billion. Alphabet raised to $195–205 billion, its second increase this year. Meta raised to $130–145 billion.

And Microsoft’s headline figure went down — to $175 billion from $190 billion.

Hold onto that one. It’s the whole story.

Subscribe now

The number they show you is the smallest one

Microsoft didn’t actually cut anything. It made an accounting change: it reclassified future data center leases from “finance leases” to “operating leases” — in plain terms, a finance lease is treated like Microsoft basically owns the building and its full cost shows up on the books right away, while an operating lease is treated more like renting, with the cost spread out and mostly kept off the main balance sheet. Microsoft also extended the assumed useful life of its buildings from 15 years to 25, which spreads whatever costs remain even thinner across more years. Same buildings, same chips, same money committed — just reported differently. The line item shrank by $15 billion; the actual commitment didn’t move an inch.

Once you notice that, you start seeing the same pattern everywhere.

Meta’s approach is a good example, worth walking through slowly rather than skipping past. Its Hyperion data center campus in Louisiana is financed through a joint venture — a separate company Meta co-owns rather than fully owns — in which an investment firm called Blue Owl holds 80% and Meta holds the remaining 20%. Blue Owl put in $23.0 billion of actual cash (equity), Meta put in $5.8 billion, and that joint venture then borrowed $27.3 billion on top of that. Meta will be the only tenant using the building. Here’s the part that matters: because Meta only owns a minority stake in the joint venture, that $27.3 billion of debt sits on the joint venture’s books, not Meta’s own balance sheet — even though Meta is the entire reason the building exists and the only company that will ever use it. The lease payments that would start reflecting this cost don’t even begin until June 2029.

Alphabet’s most recent quarterly filing (a “10-Q,” the detailed financial report public companies must file every quarter) disclosed $232.7 billion of commitments it can’t back out of — for supply contracts, content, and energy — plus $40.7 billion of future funding promised to what are called “variable interest entities” (VIEs): outside companies Alphabet has significant financial ties to and real influence over, without technically owning them outright, similar in spirit to Meta’s joint venture. On top of that, $24.1 billion of capex is sitting in “accrued liabilities” — meaning the spending has already happened, but the actual cash hasn’t gone out the door yet. That figure is double what it was the year before.

On July 24, the credit rating agency Moody’s put a number on this whole pattern across the industry. Lease commitments across these four companies total roughly $1.2 trillion, of which more than $820 billion comes from leases that haven’t even started yet — data centers still under construction. Moody’s treats these as effectively debt, since the companies are on the hook for them either way. Officially, they don’t show up as debt on the balance sheet.

I want to be clear about something here: none of this is improper. It’s all disclosed, all compliant with standard accounting rules, all sitting in footnotes and press releases rather than headline tables — which is exactly why so few people notice it, not because anyone’s hiding it. Moody’s was explicit that these four companies retain among the strongest balance sheets in the world and face no near-term risk to their credit ratings. This isn’t a fraud story, and I’d be skeptical of anyone selling it as one.

But it does mean the capex number everyone quotes is the floor of the real commitment, not the ceiling. The actual buildout is larger than the number you see reported, and the gap between the two is roughly the size of the visible number itself.

Add Microsoft’s $15 billion back on a like-for-like basis, and 2026 is closer to $750 billion in visible spending alone. The chart above was accurate the day it was published. It’s already too low — and for three years running, every single revision at every one of these companies, at every earnings report, has moved in the same direction: up.

By the companies’ own disclosures, more than half of every visible dollar goes to one line item: the chips themselves. Alphabet puts roughly 60% of its spend into servers. Microsoft says about two-thirds goes to what it calls “short-lived assets” — a polite way of saying GPUs, which wear out and get replaced far faster than a building does.

Everything else in this series is really the argument about the other half.

Spend comes fast. Getting it back comes slow.

Nearly every category of AI spending — chips, power, buildings, networking, software — started ramping up in 2022–2024. The revenue meant to justify all of it is a different story. Cloud revenue and AI-enhanced advertising are genuinely showing up now. Enterprise AI subscriptions — things like Microsoft’s Copilot or Google’s Gemini for businesses, the newest and most-hyped way this spending is supposed to pay for itself — didn’t start meaningfully contributing revenue until 2025, two to three years after the spending they’re supposed to justify.

That lag is the actual argument that just repriced four of the largest companies on earth, in nine days.

On July 22, Alphabet reported negative free cash flow of $5.9 billion — its first negative quarter since its 2004 IPO — against $44.9 billion of quarterly capex. The stock fell 7% the next day, its steepest one-day drop in years, and dragged the rest of the sector down with it. A week later, Meta posted just $784 million of free cash flow against $31.1 billion of quarterly capex, and fell nearly 10%.

Microsoft, spending a comparable amount of money over the same stretch, rose more than 14%.

Same spending, opposite market reaction, seven days apart. The market hasn’t turned against capital spending itself. It’s turned against capital spending that doesn’t come with a visible return attached.

And the return isn’t nothing — this is worth being fair about. In the same quarter Alphabet posted that negative $5.9 billion, Google Cloud grew 82% and its contracted backlog reached $514 billion. Amazon Web Services grew 37%, its fastest pace in four years. Those numbers are just as real as the spending, and they’re what the bullish case for all of this is actually built on.

So the real question isn’t whether the revenue is coming. It’s whether it arrives on the same schedule the spending has already assumed — and that schedule is set by physical things, in physical places, on their own timeline, not by how fast a company wants it to happen.

Subscribe now

Where the dollars actually land

One honest caveat before this chart: the stage-by-stage breakdown below is my own allocation model, built from industry estimates — not something any of these companies actually discloses this way. No hyperscaler reports capex broken out by “stage of the AI buildout.” The overall totals are theirs and verified; how I’ve distributed that spending across stages is my own estimate, and I’d rather say that plainly than let the confident-looking numbers imply more precision than they have. It’s built on the original $725 billion figure, before the late-July upward revisions.

Two things worth sitting with here.

First: $380 billion — more than half the total — goes to a single stage. Memory and fabrication, almost entirely GPU and HBM (high-bandwidth memory) purchases. That one number explains more about why chip and memory stocks move the way they do than any amount of narrative ever could.

Second, and less obvious: six of the twelve stages in this buildout get no dedicated capex line at all. Raw materials, neocloud leasing, cluster software, the data pipeline, model training, and deployment are all real, necessary stages — they’re just funded as ordinary operating expense, or folded quietly into the bigger numbers above, rather than broken out anywhere.

Which comes back to the same lesson as the Microsoft and Meta examples above. A stage being absent from the capex line doesn’t mean it’s absent from the buildout. It’s a disclosure choice — and it’s exactly what determines where an individual investor can actually get direct exposure to a given stage, and where they can’t.

The uncomfortable math

Three-quarters of a trillion dollars is about as much committed capital as an entire industry can realistically throw at a problem — and the real commitment is even larger than that, sitting in vehicles and leases that don’t even begin until 2029.

It still can’t buy its way past a permitting office, a transformer factory, or a grid operator’s interconnection queue. Money moves at the speed of a wire transfer. Concrete, copper, and turbines move at the speed of concrete, copper, and turbines — and no amount of capital changes that second speed.

The Idaho reactor and the Arizona packaging plant are the tell. Both exist because of this cycle. Neither delivers anything before 2027, and Arizona not until 2028. The capital arrived years ahead of the capacity to actually use it — and I think it’s going to keep arriving that way, into a physical world that simply can’t absorb it any faster than it already is.

That gap — capital moving fast, physical capacity moving slow — is where the real investment question in this whole cycle actually lives. Part 0B picks up there: the twelve physical stages this money has to survive, and which one is actually holding up all the rest.

Disclaimer

This article reflects my own research into public company filings and is provided for general informational and educational purposes only. It is not financial advice, and I am not a financial advisor. Nothing here is a recommendation to buy, sell, or hold any security mentioned. Figures are sourced from Q2 2026 earnings releases and calls (Alphabet, July 22; Microsoft and Meta, July 29; Amazon, July 30), Alphabet’s Form 10-Q, Moody’s commentary via CNBC (July 24, 2026), Meta and Blue Owl’s Hyperion joint-venture disclosures, FactSet, and capex estimates from Goldman Sachs and CreditSights. The stage-by-stage spending allocation is explicitly my own estimate, not a company disclosure, and shouldn’t be read as more precise than that. Company guidance, lease structures, and accounting treatment can all change quickly and are current only as of the date noted at the top of this piece — verify anything material directly before acting on it. This is a passion project sharing my own research process, not investment advice tailored to your situation — consult a licensed financial advisor before making any investment decision.

Subscribe now

I can’t make that assumption. There’s real uncertainty in my own life about how long we’ll stay in one place — it could be a career change, family ties pulling us elsewhere, aging parents, or simply not being sure we want to stay in this city for three decades. Whatever the reason, if there’s real uncertainty about whether you’ll be in the same house in 10 years, or 2, the real question isn’t “should I rent or buy” — it’s “if I have to leave, which way of having bought (or not bought) leaves me in the best position?”

That’s a genuinely different question, and it has more than two answers.

(I also built a free, interactive calculator for this — six scenarios, your own numbers. More on that once you’ve seen how the scenarios actually work, further down.)

The six ways this can actually play out

I modeled six real scenarios, not two:

1. Buy, pay it off before I have to move, sell it, and invest the proceeds.

2. Buy, pay it off before I have to move, keep it, and rent it out through an agency.

3. Buy, and just... stay the full 30 years. Never move.

4. Buy, still carrying the mortgage when I have to move, keep it and rent it out through an agency.

5. Buy, still carrying the mortgage when I have to move, sell midway and invest the proceeds.

6. Never buy at all — rent the entire time, and invest the difference. (Included as my reference point, not because I think it’s my best option.)

As of today, I’m targeting scenario 1 — pay it off aggressively, and if I do have to leave, sell it and invest the proceeds rather than keep it as a rental. Here’s the full comparison, and the reasoning behind it.

The real numbers, side by side

I ran all six scenarios through the calculator below using a $700,000 home, current-ish mortgage rates, and an exit/transition point on the shorter end — matching the general timeline I described in my first article (years, not decades) rather than picking a rounder but less honest assumption. Here’s what came out, ranked by after-tax final net worth after 30 years:

A few honest things this table says, in order of how much they matter:

Subscribe now

Renting throughout comes out on top, by the numbers. I’m buying anyway. That’s a lifestyle choice for my family — stability for my kid, control over the space I live in, not being subject to a landlord’s decisions — not something I’m claiming the spreadsheet supports. If pure net worth were the only goal, renting and investing the difference wins clearly in this model. I’m comfortable saying that plainly rather than reverse-engineering a reason the numbers favor buying, because they don’t.

Within the “buy” options, keeping the house as a long-distance rental (scenario 2) is my least favorite, and the math backs that up. Property tax and maintenance scale with the home’s appreciating value, while the rent you can actually charge grows more slowly. Over a couple of decades, that gap compresses the rental’s profitability — carrying costs quietly outgrow the income meant to cover them — while cash in the market just keeps compounding at its own rate the whole time.

The more interesting nuance: scenario 5 (never aggressively paying off the mortgage, just investing the spread instead) actually beats my own choice, scenario 1, by about $571,000 in this model. I’m choosing scenario 1 anyway, and I think that’s still the right call — here’s why.

Why I’m picking certainty over the higher expected number

In plain terms: the safer plan wins on “for sure,” the riskier plan wins on “on average” — and when your timeline is genuinely uncertain, “for sure” is worth paying for.

Scenario 5’s edge only exists if two things hold: the market actually delivers something close to its long-run average, and I actually invest the monthly difference every single month for years, without fail. Neither is guaranteed.

• Even a shorter runway doesn’t give the market much time for its average to reliably show up. The equity-beats-mortgage-rate case is a long-run average — over one or two decades it tends to show up, but over a stretch as short as a few years, market volatility can easily swamp it. And my own timeline is a best guess, not a floor — my actual exit could turn out to come sooner still.

• Paying down the mortgage is a forced, guaranteed return. Investing the spread requires discipline every month, indefinitely. Scenario 5’s advantage evaporates the moment the “invest the difference” habit slips for even a stretch of time — and it’s a lot easier to keep a habit that happens automatically (a fixed mortgage payment) than one that depends on me voluntarily moving money every month for years.

• The gap between scenario 1 and scenario 5 shrinks the shorter my actual timeline turns out to be. The $571,000 shown here already reflects my realistic case on the shorter end, not an optimistic longer one — and if I end up moving even sooner than that, the margin narrows further, to the point where the certainty of being debt-free is worth more to me than whatever’s left of the expected-value edge.

So my choice isn’t “I didn’t understand the math favors scenario 5.” It’s a risk-adjusted choice: scenario 5 wins on average, scenario 1 wins on certainty and on being a plan I can actually stick to without relying on my own future discipline holding up for years. Given real uncertainty about my own timeline, I think that’s worth the roughly $571,000 difference this model shows.

One important thing worth flagging if your situation looks different from mine: this entire calculation is sensitive to your actual mortgage rate, and mine assumes something close to current rates. If you locked in a mortgage during the 2020-2021 pandemic lows — plenty of people got 30-year fixed rates in the 2s or 3s — the math here looks meaningfully different, possibly different enough to flip my own conclusion. Paying down a 2.75% mortgage aggressively “costs” you a much smaller guaranteed return than paying down one at 6.3%, while the market still offers roughly the same expected 10% either way. That gap between your specific rate and the market’s expected return is the entire engine behind scenario 5’s edge — the lower your rate, the stronger the case for not rushing to pay it off, and the bigger the argument for scenario 5 (or something like it) over scenario 1. If that’s you, don’t just take my $571,000 number at face value — plug your actual rate into the calculator below and see how much that number moves. It could move by a lot.

Two things I’m not covering here, on purpose

This article is about the buy/rent/sell decision itself — it doesn’t touch what happens to this house, or anything else, if something happens to me. That’s a real gap, and not just for the money: with a young kid, questions about custody and guardianship matter as much as anything financial. I’m planning a full article on will vs. trust vs. estate plan, specifically for a family in my situation, and I think it’s genuinely necessary reading, not just a nice-to-have.

I’m also planning a separate piece on an entirely different real estate question: buying a second property specifically as a long-term rental or short-term rental (Airbnb-style) investment — different math, different goals, and a different answer than the primary-residence decision covered here.

The tax wrinkle almost nobody explains correctly

Here’s the part of this decision that a lot of generic advice gets subtly wrong, and it matters most if you’re the kind of person actually facing the “what if I leave” question — anyone who might end up living abroad down the line, for any reason: what happens tax-wise when you sell US property while living outside the US depends on your tax residency status at the time of sale — not your immigration status, and not your citizenship.

Specifically, there’s a US law called FIRPTA (Foreign Investment in Real Property Tax Act) that requires a buyer to withhold a chunk of the sale price — commonly 15% of the gross sale price, not the profit — when the seller is classified as a “foreign person” for tax purposes at the time of the sale. That withholding isn’t the final tax bill; it’s a deposit against whatever you actually owe, refunded later once you file a return. But it can tie up a meaningful amount of cash for months.

Subscribe now

The part that surprises people: whether you’re a “foreign person” here is a tax-residency question, not an immigration one. It comes down to things like the green card test and the substantial presence test — how many days you’ve spent in the US, and whether you still hold resident status — not your original immigration classification, and not your citizenship elsewhere. A green card holder who keeps their status and meets the presence test is treated like a US citizen for this purpose and isn’t subject to FIRPTA at all. Someone who leaves, lets that status lapse, and later sells while classified as a nonresident alien could be. This is exactly the kind of detail that depends heavily on your specific facts, and it’s worth confirming with a tax professional against your actual situation before assuming either outcome.

I built this uncertainty into the calculator directly, as a toggle, rather than picking one assumption and hiding the tradeoff.

Run your own numbers here

Every input above — my home price, my rent, my timeline — is specific to me. Yours won’t match. That’s the entire reason this exists as a calculator instead of a table: the six-scenario ranking above could flip completely with a different home price, a different rate environment, or a different exit year.

This matters most if your mortgage rate is nowhere near mine. If you’re sitting on a pandemic-era 2-3% rate, the calculator is the actual way to find out where you land — not my $571,000 figure, which was computed at a much higher rate than that.

→ Try the interactive calculator yourself — plug in your own home price, your own rent, your own guess at how long you’ll actually stay, and see which of the six scenarios comes out ahead for your situation, not mine.

It defaults to a $700,000 home and 30-year historical averages for everything else, but every single field is adjustable. If you take away one thing from this article, I’d rather it be “go run your own numbers” than “here’s what I did” — my answer is genuinely specific to my own timeline and risk tolerance, and yours should be too.

A few things worth knowing about how to read it:

• It shows both a pre-tax, mark-to-market figure and an after-tax, fully-liquidated figure for every scenario, since those can tell different stories — a scenario can look best on paper and still lose some of that edge once everything is actually sold and taxed.

• The FIRPTA toggle only changes the outcome if you mark that you’re relocating abroad and keeping the property as a rental — it’s modeling the specific friction of being a nonresident landlord, not asserting it always applies to everyone.

• Every default rate is labeled as a historical average, not a forecast. Change them. My own numbers will look different from yours, and that’s the point of building this as a tool rather than just publishing my own result.

Where I landed — for now

Scenario 1 is my working answer: pay it off aggressively, and sell rather than hold if I have to leave. I’m treating it as a current best guess, not a permanent commitment — if mortgage rates, market conditions, or my own timeline shift meaningfully, I’d rerun this model before assuming the conclusion still holds. That’s really the point of building it as a calculator instead of just a static number: the honest answer to “should I buy, and what do I do if I have to leave” was never going to be one number that stays true for 30 years.

If you run your own numbers through the calculator, I’d genuinely like to know what you land on — especially if you’re sitting on a much lower mortgage rate than mine, since that’s the input most likely to flip the whole answer. Drop your result in the comments, even just “scenario 1” or “scenario 5” — I’ll read all of them.

Disclaimer

This article and the accompanying calculator describe my personal reasoning and are provided for general informational and educational purposes only. Nothing here is financial, tax, legal, or immigration advice, and I am not licensed to give any of those. The calculator uses simplified assumptions (constant monthly budget, historical average returns, approximated tax treatment) that will not match your actual mortgage terms, local property tax rate, or personal tax situation — it’s a planning model, not a tax return or legal document. FIRPTA, capital gains, depreciation recapture, and primary-residence exclusion rules are genuinely complex, depend heavily on individual facts (tax residency status, treaty eligibility, time actually lived in the property, and more), and change over time; nothing in this article should be relied on as a description of your own tax treatment. Before making any decision about buying, selling, or renting out property — especially with any cross-border element — consult a qualified CPA, tax attorney, and/or immigration attorney who can evaluate your specific situation. This is a passion project sharing my own reasoning, not a recommendation that any of these six paths is right for you.

I’m in my mid-30s, married, with a young child, and there’s real uncertainty in our life about where we’ll be long-term — it could be a career change, family ties pulling us elsewhere, or simply not being sure we want to stay put for the next 30 years. I think of my own timeline as shorter than a lifetime — years, not decades. That single assumption reorders almost everything below.

Quick note before I get into it: I’m not a financial advisor, and neither is my partner. This is a personal passion project, not professional advice — I just like thinking through how to use the account types available to my household as efficiently as possible, especially around taxes, given the situation my family is in.

One decision alone — something most people treat as an afterthought — is worth a quarter million dollars over 20 years. And it’s not even my own money. Here’s the order I follow, and why.

Here’s where mine actually splits from what most guides recommend:

Two swaps, and they’re worth real money — here’s the reasoning behind each step.

Subscribe now

1. Emergency Fund — 6 months of expenses, in a High-Yield Savings Account

The verdict: not optional. It’s what keeps everything else standing.

Before I invest a dollar anywhere else, I keep 6 months of expenses somewhere safe and boring. A layoff for me isn’t just a lost paycheck — for me specifically, it can set off a much bigger chain reaction than it would for most people, but the underlying point holds regardless of the reason: this fund means I’m never forced to sell investments at a bad moment just to cover rent.

2. 401k — up to the full employer match

The verdict: free money that stops being small over time.

(Using a round $120,000 salary here as an easy example to run the math on — something clean to calculate with. The percentages and logic hold regardless of the exact figure.)

Everything above $120,000 is growth I didn’t earn and didn’t pay for. That’s the quarter-million-dollar number from the opening — from a decision that takes five minutes and costs nothing.

3. Roth IRA — maxed out, for me and my partner

The verdict: the single best account for a family that might not stay put long-term.

Most guides fund HSA before maxing Roth IRA. I do the opposite:

• If my life takes me somewhere else — for any reason — the account keeps growing and stays accessible. A 401k tied to an employer or an HSA tied to US medical bills can’t say the same.

• If something happens to me, it passes to my child — a US citizen — tax-free, with up to 10 years of continued tax-free growth before it must be touched.

• Even a non-working spouse can have one, as long as household income qualifies.

(Quick note if your income is on the higher side: Roth IRA contributions phase out above a certain income level — for 2026, roughly $153,000 MAGI single / $242,000 joint. If that’s you, a backdoor Roth IRA is a common, IRS-sanctioned way around it — more on that in my next article.)

This account is a big part of why I wanted to write about the Roth IRA in depth — more on that below.

4. Max out the 401k

The verdict: more tax-advantaged growth, once steps 1-3 are covered.

Straightforward — I contribute the IRS max here before moving to the next priority.

5. Mortgage — aggressive principal paydown

The verdict: for me, this beats HSA, given my timeline.

I constantly look for ways to cut total interest, not just the monthly payment:

• Refinancing when the math clearly favors me

• Shortening the term when possible (30-year to 15-year)

• Putting extra cash toward principal, not letting it sit

Since I think of my time in this house — and possibly this country — as finite, building equity I can walk away with matters more right now than longer-horizon tax-advantaged accounts. A family confident they’re staying 30+ years might reasonably rank this differently.

(A separate article is coming on whether owning property when your long-term plans are uncertain is even a good investment, or purely a lifestyle choice — for me, currently the latter.)

6. HSA — funded, but ranked lower than most guides suggest

The verdict: a good account, just not a priority for me.

I still contribute — it’s real tax-free money for healthcare. But it ranks last among the accounts above because HSA funds are hard to use once you’re living somewhere else, and a child can’t have their own HSA.

Subscribe now

7. 529 — low priority for me

The verdict: a good tool for the right family. I’m not sure I’m that family.

My child is young, I don’t know where they’ll go to college, and fewer international schools accept 529 funds. Unused 529 money can roll into a Roth IRA now — but only after 15 years, under strict limits, and into the child’s Roth, not mine. A nice safety net. Not a reason to prioritize it.

None of this is theoretical for me. If my circumstances change next year — for any reason — this order is what keeps my family’s future intact. That’s the whole point.

Coming soon: why the Roth IRA earns the top spot on this list — and starting with that article, I’m opening the books completely. Every trade in my Roth IRA and my partner’s, my allocation in percentages, and performance tracked head-to-head against the S&P 500 and Nasdaq 100, with the reasoning behind every pick.

But this channel isn’t just about one account. It’s about the full picture of running a household well — cutting waste, paying less interest, finding real deals on things like travel, and making deliberate trade-offs instead of default ones. The Roth IRA trades are one thread. The habits are the point.

Disclaimer

This article is for general informational purposes only and does not constitute financial, tax, legal, or immigration advice. The author is not a licensed financial advisor, CPA, or attorney. All figures are hypothetical and illustrative; past performance does not guarantee future results, and all investments carry risk of loss. Individual circumstances vary — consult a licensed professional before making financial, tax, or legal decisions. The author and this publication disclaim all liability for actions taken based on this content.

Turns out, not really — and it’s a bit more interesting than a simple “not really” once you actually look closely. Here’s all three parts of that story.

What $10,000 became

Same starting point (1972 — right after gold left the dollar standard and REIT data starts), same 54 years, wildly different endings:

Stocks beat real estate by about $375,000. Stocks beat bonds by over $1.5 million. None of this is a secret — this is just what long-run compounding at different rates actually looks like when you let it run for half a century.

Quick honesty check on “real estate” in this comparison: for most people, real estate wealth isn’t publicly-traded REIT shares — it’s the house they actually live in. I used REITs anyway, for a simple reason: there’s no clean, investable index that tracks “the average American’s primary home” the way there is for stocks. REITs are the closest thing that trades daily, reports real numbers, and lets you compare apples to apples over 50 years. Think of it less as “here’s what your house did” and more as “here’s what real estate as an investable asset class has actually returned” — a related but genuinely different question.

Why “stocks barely beat real estate” is actually hiding a bigger story

Here’s something I almost missed. The 10.0%-vs-9.5% gap in the chart above makes stocks and REITs look nearly tied over the full 54 years. But when I looked at the actual data broken out by trailing period instead of one smoothed average, the real picture is much more dramatic — and it’s a genuine reversal, not a rounding difference:

Over the full 1972-2024 stretch, and over the 40, 30, and 25-year windows, REITs actually beat the S&P 500 by a wide margin — as much as 12.6% vs. 8.0% annualized over the full period. But somewhere around the 15-20 year mark, that flips hard. Over the last 10, 5, and 1 years, the S&P has dominated instead — 23.2% vs. 4.9% in 2024 alone.

That explains why my first chart looked closer than the long-run history actually was: the smoothed “since 1972” figure blends two completely different eras together — decades where real estate quietly won, followed by a recent stretch where stocks have won emphatically. The recent S&P run has been strong enough to pull the whole-period average up toward “close to REITs,” even though for most of the last 50 years, REITs were the clear long-run winner.

Subscribe now

How people actually invested, over the same stretch

The verdict: household behavior has only matched what the numbers say three times, ever — each time right at a euphoric peak, not calmly along the way.

Real estate was the dominant American asset for most of the last century — it made sense; a home is the first real asset most families ever hold. Stocks have pulled ahead of it before, though less often than most people think: once in the late 1960s, again at the dot-com peak in 2000, and now — a third time in roughly six decades, not the second, as I’d originally had this.

Here’s the part worth being careful with. Both prior instances were followed by genuinely rough, extended stretches for stocks — the late 1960s crossover preceded a decade of poor real returns, and the 2000 crossover preceded the dot-com crash. It would be easy to read that as “this always happens right before a crash.” I don’t think that’s a fair conclusion to draw. Two prior instances is nowhere near enough data to call this a reliable pattern — it’s a genuinely small sample, and treating it as a forecasting tool is exactly the kind of overconfident prediction I don’t want to make. What I think is fair to say: it’s a real, historically rare situation, worth being aware of, not a signal telling you what happens next.

Worth adding one thing that makes this round of the shift a little different from the first two: this time, it’s arguably not pure euphoria. The trend chart above shows the S&P genuinely has crushed real estate over the last decade specifically — so household money moving toward stocks right now is, at least partly, just following where the recent returns actually were. That doesn’t make it risk-free (the same was true in 2000, right before it wasn’t true anymore), but it’s a more rational-looking shift than the dot-com version, not an identical replay of it.

One honest caveat worth naming: this kind of household wealth data is always released on a lag — even the newest breakdown available isn’t literally real-time, it’s typically a couple of quarters behind. So “recent” here means the latest solid figure, not necessarily where things stand at this exact moment.

One more thing worth clarifying, since it’s easy to misread: the third category in the chart isn’t just “bonds and cash sitting idle.” It’s bank deposits, pension entitlements, private (non-publicly-traded) business equity, and actual bonds, combined — a genuinely mixed bucket, not one big pile of cash on the sidelines.

The part worth sitting with

I don’t think this means “stocks are now overheated, sell everything” — that’s exactly the kind of prediction I don’t want to make. But it is worth noticing: every time household stock allocation has caught up to what the return data has said for decades, it’s happened at a moment of unusual enthusiasm, not steady conviction. That’s a pattern worth being aware of about your own behavior, not just the market’s.

I don’t have a clean answer for what that means for right now. I just think it’s a more useful question than most people are asking.

Long-run historical averages, not a year-by-year backtest — actual annual returns for stocks and gold especially are far more volatile than a smooth compounding line shows. Real estate proxied by REITs since no single “real estate” index exists — REIT returns and primary-home returns are related but genuinely different things. REIT vs. S&P 500 trailing-period data from Nareit/FTSE NAREIT, most recent available as of time of writing (mid-2026), not real-time. Household allocation data from Federal Reserve Flow of Funds / Distributional Financial Accounts, snapshot years shown, not a continuous series. Nothing here is a recommendation to buy, sell, or hold anything — just three charts I think are worth looking at together.

This December, I’m taking my family of 3 to the Big Island of Hawaii — round-trip flights and 5 nights at a resort, booked and paid for already, for a combined $1,014. That figure covers flights and the hotel only — not food, activities, car rental, or anything else once we’re actually there. Here’s exactly how I got those two pieces down to that number.

The rule I use before booking anything

The verdict: only use miles if you’re clearly getting more than cash would.

Every time I book travel, I check cash price and miles price side by side, for every reasonable airport option. I’m on the East Coast, and both Newark (EWR, a United hub) and Philadelphia (PHL, an American hub) are within reach, so I check both airports, both ways to pay, before deciding anything.

My rule of thumb: I’ll consider a miles booking if I’m getting over 1.5 cents of value per mile, and I’ll go for it without hesitation above 2 cents. Below that, cash is usually the better move. For this trip, both the flights and the hotel cleared 2 cents — that’s why I booked both with points.

Flights: $596.10 total, for 3 people, round trip

The verdict: the best deal wasn’t with the airline flying out of my home airport — it was found through a completely different loyalty program.

I found the best value flying American Airlines out of Philadelphia, but not by booking through American directly — I booked through Finnair, using Avios.

Here’s why that works: Finnair, American Airlines, and British Airways are all part of the Oneworld alliance, and Finnair, British Airways, Iberia, Aer Lingus, and Qatar Airways all share the same currency — Avios. That means Avios earned or transferred into any one of those programs can move freely between the others. Finnair itself isn’t a direct Bilt transfer partner, but British Airways is — so the actual path was:

Bilt points → British Airways Avios (with a transfer bonus) → moved over to Finnair → booked on American Airlines through Finnair.

Subscribe now

Finnair’s award chart prices a one-way East Coast–to–Hawaii flight on American Airlines at a flat 15,000 Avios, regardless of the exact route or connection — plus about $5.60 in taxes and fees. Alaska Airlines is nominally cheaper on paper at 13,500 Avios each way — but because of the 60% transfer bonus I had that day, I only needed to transfer 9,375 Bilt points to cover the 15,000 Avios for American. That’s actually fewer Bilt points spent than Alaska would have cost me at 13,500 points with no bonus at all. The bonus, in other words, didn’t just make American competitive with Alaska — it made American the better deal, even though Alaska’s sticker price is lower.

On May 1st, Bilt ran a Rent Day promotion giving Gold-status members that 60% bonus on points transferred to British Airways. That meant instead of transferring the full 15,000 points needed, I only needed to transfer:

15,000 ÷ 1.6 = 9,375 Bilt points to cover one seat, one-way.

For 3 people, round trip, that’s 6 one-way seats total:

• 9,375 points × 6 = 56,250 Bilt points (valued at 1 cent each = $562.50)

• $5.60 × 6 = $33.60 in taxes and fees

• Total: $596.10

American’s Main Cabin economy fare on this route includes a personal item and a full carry-on bag — no extra baggage fees to factor in.

The hard part isn’t the math — it’s finding the flight. Finnair’s own site is genuinely difficult to search for American-operated award space, since not every American flight is released for partner bookings. To get around this, I cross-check availability using two other Oneworld programs with better search tools — Alaska Airlines and Qantas — since they show the same underlying American Airlines award inventory. Once I spot a flight that looks available, I call Finnair customer service directly to actually book it, since American Airlines partner awards can’t be booked online through Finnair — only by phone or chat.

I had the exact flight identified before the May 1st transfer bonus even showed up. That’s the part that actually takes effort — the promo just made a flight I already wanted to book meaningfully cheaper.

Cash vs. miles, side by side, for my actual trip window (first half of December, 5-night gap) — flights and hotel only:

Cash price (approx.)What I paidSavingsFlights (3 people, round trip)~$2,145*$596.10~72%Hotel (5 nights)~$1,400$417~70%Total~$3,545$1,014~71%

*Cash estimate based on a comparable PHL–Kona, one-week, economy round-trip fare (~$715/person) pulled around the same December window — actual cash pricing shifts by day and by how far in advance you book, so treat this as a representative comparison, not the exact fare available the day I booked.

That works out to somewhere north of 3.5 cents of value per Bilt point on the flights alone — well past my 2-cent threshold, which is exactly why I booked this with miles instead of cash.

Hotel: $417 for 5 nights at Paniolo Greens (Wyndham)

The verdict: an Airbnb-style resort stay, paid for almost entirely in points, with free cancellation built in.

I found Paniolo Greens, a Wyndham resort on the Big Island that gives me the kitchen and space I look for on trips like this. Cash price: roughly $200/night, ~$1,400 total with taxes for 5 nights. Points price: 15,000 Wyndham points per night, taxes included — and free cancellation right up until the trip.

To fund the Wyndham points, I again used a Bilt transfer bonus — this time on April 1st, Bilt offered Gold members a 75% bonus transferring to Wyndham. I paid $125 in Bilt Cash to bump that up further, from a 75% bonus to a 100% bonus (2x) — pricing on that kind of upgrade varies by promotion, but that’s what it cost me this time.

At a 2x transfer rate, reaching the 75,000 Wyndham points needed for 5 nights took:

75,000 ÷ 2 = 37,500 Bilt points

Here’s how I think about the cost of that:

• 37,500 Bilt points, valued at 1 cent each = $375

• The $125 in Bilt Cash I spent to upgrade the bonus — I don’t value Bilt Cash at full face value here, because Bilt Cash used this way effectively converts at a $30-to-1,000-points rate. At my own 1-cent-per-point valuation, that works out to roughly 0.33 cents of real value per $1 of Bilt Cash spent — so I treat that $125 as really costing about $41.67 in equivalent value, not $125.

• $375 + $41.67 ≈ $417

Either way you value it, I turned a ~$1,400 cash stay into a ~$417 one, with the flexibility to cancel free if plans change.

Subscribe now

A couple of past trips, same approach

I’ve used this same cash-vs-miles comparison method before:

• December 2024 — Newark to Maui, via Honolulu, round trip: $558.60 for the flights, booked using Turkish Airlines miles on United. The pricing on this one actually shifted mid-search: until February 2024, Turkish priced these United award tickets to Hawaii at 7,500 miles + $5.60 one-way, and I locked in 3 one-way tickets (my outbound leg, for the whole family) at that rate. Turkish then repriced the route to 10,000 miles one-way — so my return leg cost more, 3 tickets at 10,000 miles + $5.60 each. Total: (7,500 × 3 + 10,000 × 3) miles, valued at 1 cent each = $525, plus $33.60 in taxes across 6 segments = $558.60. Turkish has since changed its Hawaii pricing again, and this specific deal no longer exists.

• May 2024 — JFK to Barcelona to Paris to Lisbon: about $700 per person, using a mix of miles and cash across Level, Air France, Vueling, and United.

I like traveling. I just don’t like paying full price for it if I don’t have to — and I’d rather spend the effort finding a good flight than spend the money not finding one.

Disclaimer

This article describes the personal travel-booking experience of the author and is provided for general informational and entertainment purposes only. It is not professional travel advice, and the author is not a certified travel agent or advisor. Award availability, transfer bonuses, redemption rates, and program terms mentioned here (including with Bilt Rewards, Finnair Plus, American Airlines, Wyndham Rewards, and any other loyalty program referenced) change frequently and without notice, and specific deals described may no longer be available or may require different amounts of points by the time you read this. Pricing, availability, and terms should always be independently verified directly with the relevant airline, hotel, or loyalty program before booking. Nothing in this article should be relied upon as a guarantee of similar results, and the author and this publication accept no liability for bookings, losses, or issues arising from following any strategy described here. This is a passion project reflecting how I personally choose to travel — not a recommendation that this approach is right for your situation.

Quick note before I get into it: this isn’t sponsored — Bilt didn’t pay me or ask me to write this. I do include a referral link further down, and if you apply through it and get approved, I may earn a referral bonus at no extra cost to you. Everything else here is my honest, unpaid take, limitations included.

The welcome bonus — worth mentioning before anything else

The verdict: this alone can cover most of the first year’s fee, before any of the ongoing math below even applies.

New Palladium cardholders can currently earn 50,000 Bilt Points, Bilt Gold status, and $300 in Bilt Cash after spending $4,000 on non-housing purchases in the first 90 days. Independent point valuations put 50,000 Bilt points somewhere around $1,000+ on their own — combined with the $300 in Bilt Cash, that’s a meaningful chunk of value sitting in year one alone, well before any of the ongoing per-dollar math below. Offers like this change over time, so verify the current terms directly with Bilt before applying — but as of when I’m writing this, it’s a real, active offer, not a promotional gimmick.

The points math — why I think it earns more than it looks like

The verdict: on paper it’s 2X. In practice, I get closer to 3.33%.

Palladium earns a flat 2X points on almost everything you put on the card. One important exception: tax payments earn zero points. I learned that one the hard way, paying property taxes on the card expecting points and getting none. Don’t make that mistake — it’s excluded entirely, not just at a lower rate.

A flat 2X isn’t unique — cards like Capital One Venture X and Wells Fargo Active Cash also land around that range. Where I think Palladium pulls ahead is the layer underneath the 2X: you also earn 4% back in Bilt Cash on the same spend, and Bilt Cash has real uses across Bilt’s ecosystem. The one I lean on most is unlocking points on rent or mortgage payments — every $30 of Bilt Cash redeemed unlocks 1,000 points on your housing payment. That’s where the extra value comes from.

A concrete example:

• Spend $100 on the card

• Base: 2X points = 200 points

• Plus: 4% Bilt Cash = $4

• That $4 in Bilt Cash unlocks points on $133 of your rent or mortgage payment (at $30 Bilt Cash → 1,000 points), earning another 133 points

• Total: 333 points on $100 of spend — an effective rate of about 3.33%
  

  Subscribe now

That’s the floor, not the ceiling. If you’re more active in the Bilt ecosystem and find other uses for the 4% Bilt Cash, you can likely do better than this. I keep it simple — most of mine goes toward the mortgage, except twice a year, when I redeem it toward a hotel stay instead (more on that below).

Two things worth knowing that I don’t personally use, but that fit some situations better than mine:

• A simpler alternative path exists — and it’s not really a tradeoff on everyday spend, just on housing. The 2X on everyday purchases is a fixed property of the card either way; it doesn’t change based on which option you pick. What actually changes is the housing side: Bilt also offers a “Housing-only rewards” option, which drops the 4% Bilt Cash entirely but automatically earns up to 1.25X points on rent or mortgage — tiered by how much of your housing payment you match in everyday spend that same billing cycle (25% match = 0.5X, 50% = 0.75X, 75% = 1X, 100%+ = 1.25X). That ceiling is actually higher than the 1X cap you can reach by redeeming Bilt Cash under the Flexible path I use. The real question isn’t “which earns more on everyday spend” — it’s whether a slightly better housing rate is worth giving up the 4% Bilt Cash and everything else it unlocks.

• A points accelerator exists for Bilt Cash too. Instead of unlocking housing points, you can redeem $200 of Bilt Cash to earn an extra point per dollar on your next $5,000 of eligible spend — effectively 3X instead of 2X — repeatable up to 5 times a year. If you don’t have rent or a mortgage payment to route through Bilt, or you’ve already maxed out what you want to put toward housing, this is the other place that 4% Bilt Cash can go.

One thing worth adding since it’s not in the math above: Palladium also gives you a separate $200 annual Bilt Cash bonus, credited automatically each January, on top of the ongoing 4%. I don’t count that in my fee math below — it’s just extra.

The $495 fee — how I offset it

The verdict: easily, between the hotel credit and the extra 1.33%.

I don’t like paying fees I can’t clearly offset, so here’s how I think about this one.

Palladium gives $400 in hotel credit per year, split into two $200 credits for bookings through the Bilt Travel portal (minimum 2-night stay each). I do use these, but since you could sometimes find the same room cheaper elsewhere, I value this credit conservatively — at 65% of face value, not 100%. That’s $260 of real value toward the fee.

$495 − $260 = $235 left to offset. That’s easily covered by the extra 1.33% from the Bilt Cash-to-housing-points mechanic described above. On $1,500/month of spend on this card (my primary card, so that’s realistic) — $18,000/year — 1.33% works out to about $239. That alone clears the remaining fee in any ongoing year. And that’s before counting two more things: the separate $200 annual Bilt Cash bonus mentioned above, and a milestone bonus every Bilt member gets — $50 in Bilt Cash for every 25,000 Bilt Points earned toward status. At that same $18,000/year of spend, my combined points (the base 2X plus what gets unlocked on housing) land around 60,000 points a year — enough to clear two of those 25,000-point milestones, for another $100 in Bilt Cash. Stack all of it together, and the $495 fee isn’t a close call for me in an ongoing year — and in year one specifically, the welcome bonus alone likely covers most or all of it before any of this other math is even needed.

One more thing worth knowing: on the 1st of every month (”Rent Day”), Palladium doubles your points rate from 2X to 4X on everyday spend, up to 1,000 bonus points. If you know a big purchase is coming, timing it for the 1st is free extra value.

Redeeming the hotel credits — how I’ve made it work

The verdict: plan ahead and you can stretch two credits into one longer trip — just don’t overcommit to a trip you might cancel.

If you have a trip planned more than 6 months out and it’s longer than 4 nights, you can actually combine both semiannual $200 credits into a single stay — but it requires making two separate reservations, and if there’s real chance you’d cancel, I wouldn’t recommend booking that far ahead. You’ll lose the earlier credit if it expires before you use it (these are on a rolling ~6-month window from when they’re issued). For the actual booking, I typically pair the $200 credit with up to $100 of my own accumulated Bilt Cash to cover a 2-night stay.

Transfer partners — where I think the card’s real value lives

The verdict: this is the best part of the card, and it’s not close.

Bilt has more transfer partners than most premium cards — more than 20, compared to what you’d get with Chase Sapphire Reserve or Amex Platinum. But the count isn’t the whole story; the quality matters more. Of the partners available, I rate World of Hyatt and Alaska’s program (now branded Atmos Rewards) as the two best for getting outsized redemption value. Chase also has Hyatt — but none of the major competing cards have Alaska/Atmos. I’ll do a separate article on exactly why I rate these two so highly.

One important catch: Bilt’s best transfer bonuses are almost always one-day-only promotions, on the 1st of the month. You need to be ready to act immediately, not “get to it this week.” This is exactly the mechanic that got me over 2 cents per point of value on my Hawaii trip (see my earlier travel article) — it only worked because I was ready the moment the bonus posted.

Other benefits worth knowing about

The short version: strong lounge access for the whole family, a genuinely useful Costco workaround, and one thing I’m still waiting on.

Lounge access. Palladium includes Priority Pass — not just for you, but for 2 additional guests, which fits my family of 3 exactly. If you need more guest passes than that, you can buy extra ones with Bilt Cash at a discount — I personally value Bilt Cash at roughly 0.33 cents per dollar of face value in this kind of use, based on what it’s actually worth to me elsewhere in the ecosystem. This guest allowance is a real differentiator: only true premium cards like Sapphire Reserve and Amex Platinum extend lounge access to guests this generously — even a mid-tier card like Venture X only covers the primary cardholder. One caveat: Bilt has no lounges of its own, unlike those three cards, though Priority Pass does get you into some Chase and Capital One lounges in certain cases.

The Costco trick. Costco warehouses and gas stations only accept Visa in the US, and Bilt is a Mastercard — so it doesn’t work directly in-store. My workaround: buy a Costco Cash Card online (the e-gift-card version posts almost instantly), load it into the Costco app wallet, and use that in-store or at the pump. Simple, and it lets me keep earning Bilt points on Costco spend that would otherwise be locked out.

Subscribe now

What I’d still like to see. Palladium doesn’t currently include primary rental car insurance, which is why I still keep a Capital One Venture X in my wallet specifically for that. Why I still hold onto Venture X for that one benefit is a topic for another day.

If you want to check it out

If you’ve read this far and you’re considering the card, here’s my referral link: bilt.com/card

Full disclosure: this is a referral link, and if you apply and get approved through it, I may earn a referral bonus — at no extra cost to you. You’re obviously welcome to go directly to bilt.com instead if you’d rather not use it; nothing in this article changes either way.

Disclaimer

This article reflects my personal experience with and opinion of the Bilt Palladium Card. It is not sponsored by Bilt or any other company, and no one paid me or reviewed this content before publication. It does include a referral link, disclosed above, through which I may earn a referral bonus if you apply and are approved — this creates a financial incentive on my part, which I’m disclosing so you can weigh it accordingly. It is not financial advice, and I am not a financial or credit advisor — card terms, benefits, transfer partners, bonus categories, and fees change frequently and without notice, so verify current terms directly with Bilt before applying for or relying on anything described here. Individual results depend on your own spending patterns, credit profile, and how you use the card’s benefits; your value from this card may differ meaningfully from mine. This is a passion project sharing what’s worked for my household, not a recommendation that this card is right for your situation.