Amazon's opportunity is increasingly defined by the economics of enabling commerce and AI at scale. Its platforms can supply valuable capacity and reach, but the same layers — infrastructure, fulfillment, fees, and compliance — can shift cost and operating risk onto the customers who use them. That tension matters because durable competitive advantage depends on two things at once: superior capability, and a total cost of participation that remains transparent and manageable for sellers and cloud users. A road network only creates value when the tolls, permits, and maintenance regimes still let traffic flow.
The analysis below works through three layers of that system in order: the AI capacity AWS is building, the billing and governance rules that decide whether that capacity earns its keep, and the seller-side cost stack in the marketplace, where India's festive season provides the sharpest current test.
AWS: Heavy Roadbed, Demanding Traffic
What is being built
The product direction on AWS is toward exceptionally dense AI infrastructure. EC2 UltraClusters can scale P6e and P6 systems to tens of thousands of GPUs on petabit-scale nonblocking networks 12, and at the top of the stack a P6e-GB200 UltraServer combines 72 GPUs, 13,320 GB of HBM3e memory, 405 TB of storage, and 28,800 Gbps of EFA networking 12. This is heavy-duty roadbed, engineered for aggregate loads that would have been impractical a few product cycles ago.
Why utilization, not scale, settles the economics
Two demand-side currents favor this build-out. Inference is moving from episodic to recurring utilization 2, and it is increasingly bound by memory and bandwidth rather than raw compute 2. Capacity that runs continuously against steady traffic is the pattern that pays for dedicated accelerated hardware.
But traffic on a new highway is never guaranteed. Demand can remain concentrated in training periods, leaving capacity underused at other times 1, and electricity availability constrains both global expansion and inference at the 100,000-card scale 2. The reasonable conclusion is that AWS's high-performance positioning creates an execution requirement as much as a hardware advantage: customers need workloads that are sufficiently sustained and operationally mature to absorb the cost of dedicated accelerated capacity. A bridge earns its keep only if traffic crosses it.
Billing and Governance: Where Unit Economics Are Actually Decided
Standing resources versus jobs
Cost visibility and governance are central to that execution requirement. AI billing distinguishes standing resources, which are billed while they exist, from jobs, which are billed for runtime 11. Provisioned throughput and notebook instances are standing resources 11, while batch inference is a job resource 11. The distinction is economic rather than academic: idle always-on capacity is a meter that runs whether or not the plant is producing.
The monitoring gap
That risk is compounded by a blind spot in the gauges. CPU-centric monitoring can leave idle GPU-bearing workloads invisible to FinOps and infrastructure teams 3; if the instrumentation measures axle load while the problem is engine temperature, the failure arrives unannounced.
Security as a cost multiplier
Security weaknesses can compound these costs rather than merely accompany them. Internet-facing database endpoints receive continuous scanning and connection attempts that consume connections and CPU 10, and compromised cloud accounts are most often monetized through compute abuse, including mining on large or GPU instances 10. The strategic implication is that AWS customers increasingly require security, identity, utilization, and cost controls to operate as one discipline rather than as separate administrative functions. On a physical road network, the tolling system, the weigh stations, and the traffic cameras are not run in isolation; the same logic applies to cloud operations.
The Marketplace: Landed Cost Beyond the Sticker Price
What FBA sellers actually pay
The marketplace and fulfillment side reflects a comparable total-cost challenge. For FBA sellers, landed cost extends beyond product price to freight, duty, preparation, and Amazon fees 8, while referral fees, supplier shipping, warehousing, and logistics are further costs sellers may overlook 9. The invoice price is one layer of pavement on a much deeper roadbed.
Fixed obligations at the border
Cross-border expansion adds fixed obligations that do not scale with sales. Storing inventory locally triggers country-specific VAT registration regardless of revenue 13, with professional registration and compliance estimated at €1,200–€3,000 per country annually 13. Suggested economic thresholds — cross-border fulfillment below roughly €40,000 of annual country sales, local storage above €80,000–€100,000 — are explicitly the author's unaudited notes based on peer experience, useful as decision heuristics rather than established benchmarks 13. The provenance deserves flagging: a rule of thumb and a certified load rating are different instruments, and they should be labeled accordingly.
India's Festive Squeeze: Peak Demand Meets Rising Tolls
Fee and logistics changes arrive together
The issue is particularly acute in India ahead of the festive season. Amazon and Flipkart revised seller fee and penalty structures before festive shopping — commercial platform-policy changes rather than government mandates 4,5. In the same window, Delhivery's acquisition of Ecom Express increased its parcel volumes and pricing leverage with D2C brands 7, while shipping already represents 15%–25% of order value for brands such as boAt and Mamaearth 6.
Peak demand does not hold prices down
Diwali drives substantial logistics order spikes 6, but peak demand does not prevent price increases 7. With simultaneous fee hikes by the three major logistics providers leaving D2C sellers with limited alternatives 6, higher shipping costs are expected to be passed through to consumers 7. This supports an interpretation that Amazon's seller ecosystem faces a near-term trade-off between preserving transaction volume during a high-demand period and protecting merchant margins and consumer price competitiveness. Raising tolls when the road is busiest moves traffic: some of it finds another route, and some of it stays home.
What the Evidence Supports — and What It Does Not
The underlying pattern
The broader pattern is not simply that Amazon can charge for scarce infrastructure or ecosystem access. Its position is strongest when its services reduce complexity enough to justify their direct and indirect costs. In cloud, that means translating leading-edge AI capacity into reliable utilization and disciplined governance. In commerce, it means ensuring that fulfillment, fees, and regulatory requirements do not make growth uneconomic for smaller sellers.
What remains uncertain
The material does not establish how Amazon will price, subsidize, or otherwise manage these trade-offs, so the financial effect remains uncertain. Three measurable items will indicate the direction: whether AI workloads sustain utilization beyond training peaks; whether seller landed costs — fees, freight, and VAT — stay proportionate to order economics in key markets; and whether India's festive pass-through compresses merchant margins or is simply absorbed by consumers. None of these outcomes is settled by the evidence above, and each will show up in utilization, take-rate, and price data well before it appears in strategy documents.