--- title: "The $200 Billion AI Tax on Your Cloud Bill" url: "https://bitrefinery.com/blog/200-billion-ai-tax-cloud-bill" description: "Hyperscalers are spending hundreds of billions on AI infrastructure — and they're passing that cost straight to you through repriced compute, storage, and egress. Here's what's actually happening and what you can do about it." author: "Bit Refinery Infrastructure Team" date: "2026-04-12" lastmod: "2026-04-12" tags: ["cloud costs", "bare metal", "infrastructure", "ai infrastructure", "egress fees", "clickhouse", "gpu hosting", "cost optimization"] source: "blog CMS" --- # The $200 Billion AI Tax on Your Cloud Bill AWS, Azure, and Google Cloud are going to spend somewhere north of $200 billion on infrastructure in 2025. That's not a typo. Microsoft alone is dropping $80 billion. Google's in for $75 billion. Amazon's somewhere around $100 billion when you add it all up. And here's the thing nobody's saying out loud: *you're paying for it.* Not through some direct line item. It's subtler than that. It shows up in the way traditional compute pricing has quietly stopped getting cheaper. It shows up in egress fees that haven't budged in years despite bandwidth costs dropping dramatically. It shows up in reserved instance pricing that used to be a great deal and now... isn't. Call it the AI tax. It's real, it's growing, and most engineering teams haven't fully internalized what it means for their infrastructure budgets. ## What's Actually Happening For about fifteen years, cloud pricing followed a pretty reliable pattern — Moore's Law-adjacent deflation. Every year or two, the hyperscalers would drop prices on compute and storage, and everyone kind of assumed that would just... keep happening. It hasn't. Not really. Not since around 2021. The reason is straightforward: the hyperscalers are in a race. Not for your workload specifically — for AI dominance. They're building massive GPU clusters, custom silicon (TPUs, Trainium, Maia), ultra-high-bandwidth networking between racks, and the power infrastructure to run all of it. Tens of gigawatts of new data center capacity, globally. [Your AI bill is subsidized](/blog/ai-subsidy-cliff) for now, but that capital has to get recovered somewhere. And the place it gets recovered is across the entire customer base — including the boring, non-AI workloads that are just running databases, serving APIs, and storing data. Your Postgres instance isn't subsidizing your ML team. It's subsidizing *their* ML ambitions. ## The Numbers Are Pretty Stark Let's make this concrete. Take a memory-optimized workload — something like a ClickHouse cluster or a heavy analytics database. On AWS, an `r6i.metal` instance with comparable storage runs around **$10,658/month**. That's before egress. If you're moving data — and any serious analytics workload is moving a lot of data — you're looking at $0.08 to $0.09 per GB out of AWS. For a team pulling 200 TB/month, that's another **$16,200 in egress fees alone**. Every month. Just for the privilege of accessing your own data. A comparable bare-metal configuration — 80 cores, 1 TB RAM, 44 TB SSD storage — runs **$2,800/month** at Bit Refinery. Egress? Zero. Unlimited bandwidth included. The math isn't close. And it's getting less close every year as hyperscaler capex climbs. ![Cost comparison chart showing AWS vs Bit Refinery for compute and egress fees](/api/storage/files/blog-images/infographic-1775991677969.png) ## "But We Need the Flexibility" This is the argument I hear most often, and it's not wrong — it's just incomplete. Yes, cloud flexibility is real. Spinning up 500 instances for a one-week batch job and then spinning them back down is genuinely useful. Nobody should be arguing against that. But most workloads aren't like that. Most workloads are baseline — they run 24/7, they have predictable resource requirements, and they've been running that way for years. Your data warehouse isn't elastic. Your analytics cluster isn't elastic. Your ML training jobs might spike, but the storage and preprocessing pipeline underneath them sure isn't. For those workloads, you're paying the flexibility premium for flexibility you're not using. And now you're *also* paying the AI infrastructure tax on top of it. The smarter model is to [baseline on bare metal and burst to public cloud](/blog/hybrid-cloud-bare-metal-baseline-burst-public-cloud) for genuine spikes. You get predictable costs where predictability matters and real flexibility where it actually helps. ## Egress Fees Are the Sneakiest Part I want to dwell on egress for a second because I think it's the most underappreciated part of this whole dynamic. The actual cost of moving data across a network has been falling for years. Bandwidth is cheap. Transit is cheap. The $0.08/GB egress fee AWS charges has almost nothing to do with their actual cost of moving bits — it's a retention mechanism. Make it expensive to leave, and customers stay. What's wild is that this fee structure also makes certain architectures genuinely worse. Want to run a hybrid setup with some workloads on-prem and some in cloud? Every time those systems talk to each other, you're paying. Want to use multiple clouds for resilience? Same problem. The egress fee is a tax on good architecture. Some providers are starting to push back on this. Bit Refinery's Denver facility includes a free Google Cloud Interconnect — dedicated private peering with $0 egress from bare metal to GCP and dramatically reduced egress coming back the other way. This type of [free private peering](/blog/why-we-give-away-cloud-interconnect-fees) makes hybrid architectures actually economical instead of theoretically nice. ## What Smart Engineering Teams Are Doing The teams that are getting ahead of this aren't abandoning cloud entirely — that's not realistic for most organizations. They're being more deliberate about what runs where. Specifically: **Moving stable, high-memory workloads to bare metal.** ClickHouse, Trino, large Postgres instances, ML training pipelines — anything that needs consistent resources and generates a lot of data movement. The economics are dramatically better off cloud for these. **Treating egress as a first-class cost.** Auditing where data moves, why it moves, and what it costs. A lot of teams genuinely don't know their monthly egress bill until someone digs into it, especially since [GCP recently doubled egress rates](/blog/gcp-doubled-egress-rates-free-private-peering-hybrid-architecture) for certain types of traffic. **Rethinking reserved instances.** Three-year reserved instance commitments made a lot of sense when cloud was the only option. With bare metal providers offering month-to-month pricing at a fraction of the cost, locking into a multi-year cloud commitment looks a lot less attractive. **Bringing GPU workloads in-house.** Cloud GPU pricing is brutal — H100s on AWS run $30+/hour. If you're doing any meaningful volume of training, owning or colocating your own GPUs and paying for managed infrastructure is almost always cheaper. Bit Refinery's BYOGPU program, for instance, starts at $600/month per GPU — compared to $20,000+ you'd spend renting the equivalent on a major cloud for the same period. ## The Uncomfortable Truth The hyperscalers made a bet that AI would be the next platform shift — and they're probably right. But they made that bet with capital that has to be recovered, and the recovery mechanism is your cloud bill. This isn't a conspiracy, it's just economics. When you spend $200 billion on infrastructure, you need a return on that investment. The return comes from customers who haven't done the math on alternatives. Do the math. The alternatives are better than they've ever been — better hardware, better managed services, better interconnects, better SLAs. The AI boom is a great time to be building AI products. It's a terrible time to be naively running everything on hyperscaler infrastructure without questioning whether you need to. Your cloud bill is going up. The question is whether that's because your business is growing, or because you're quietly funding someone else's ambitions. --- *Bit Refinery provides bare metal hosting, GPU colocation, and managed data services from our Denver and Seattle data centers — with $0 egress fees and predictable monthly pricing. [Talk to our team](https://bitrefinery.com/contact) if you want to run the numbers on your specific workload.*