The archive · Page 4 of 6
The serving stack, network fabric, and cloud-account substrate beneath production AI, where every throughput claim collides with rebuild windows, egress invoices, and control-plane risk.
73–96 of 130 articles · Newest first
Fixed-corpus RAG collapses re-chunking, freshness sync, and citation grounding. The Bible's book.chapter.verse scheme isolates churn-driven complexity from the algorithmic.
Vercel's KIKO Milano case study has no QPS, cost, or cold-start data. Collapsing skips checkout paths, Fluid savings assume I/O-bound work. Run the numbers before the peak.
Vercel Postgres is a Neon resell now owned by Databricks. The bundled database trims setup friction for Next.js teams but stacks three vendors where most assume one.
The 20B RLHF on 24GB claim is real but the VRAM cost is the duplicated reward model, not the LoRA adapters. Batch size, sequence length, and throughput pay the price.
Vercel's Flat Rate CDN beta replaces per-GB egress with a fixed fee; spiky workloads win, steady low-traffic sites pay more, and the Cloudflare-fronting cost case weakens.
DeepSeek hit 17% of tokens on Vercel's AI Gateway in May 2026, but no source confirms where the weights run. Jurisdiction, not model origin, is the new gate.
Vercel markets Next.js and the AI SDK as open source and portable, but the paid platform, from deploy previews to Fluid Compute, does not travel with the code.
Vercel's January 2026 Vercel-Cache-Tag ship brings surrogate-key cache invalidation to every plan, moving the cache contract into application-owned tag strings.
Vercel AI Gateway reportedly routes GLM 5.2 Fast through Wafer, putting Zhipu's 754B-parameter MoE coding model one config entry from any Vercel app. Routing is unverified.
Vercel's tag-based cache invalidation shifts cost from each purge call to per-response edge metadata, forcing teams to design a low-cardinality tag taxonomy up front.
Prisma Postgres now bills through Vercel Marketplace, turning the ORM company into your database vendor. Transactional pool limits constrain serverless workloads.
OpenAI on AWS Bedrock is unconfirmed, but a loosened Microsoft deal would let AWS shops draw down committed spend rather than pay cross-cloud egress for GPT-class models.
Vercel Observability surfaces per-function cost and error metrics natively, retiring basic Datadog for cost attribution, but cannot trace cross-service latency chains.
Vercel's AWS Databases marketplace streamlines IAM and billing but leaves cross-region latency and pooling on the buyer. The 64ms benchmark holds only on greenfield accounts.
Turso on the Vercel Marketplace embeds SQLite in serverless functions, eliminating the read-side connection pool at the cost of write conflict aborts and replica lag.
Vercel has been on the AWS Marketplace since 2024. The real shift is that an EDP commit can absorb the spend, coupling the frontend host to AWS and deepening the lock-in.
CrossPool splits MoE FFN weights and KV-cache into separate GPU pools and ships hidden states across the boundary, trading VRAM residency for an interconnect bottleneck.
Vercel's in-function concurrency lets one warm instance serve many requests. Cold starts and idle-I/O bills drop, but CPU-bound handlers contend and shared state can race.
CAREATTACK (June 2026) edits RAG retriever weights to surface attacker-chosen passages, bypassing corpus defenses and moving the trust boundary onto the embedding model.
Vercel lifted its CDN origin timeout from 30 to 120 seconds, relocating rather than removing a failure boundary and holding wedged connections four times longer before a 504.
Gradio-Lite runs the Gradio Python runtime as WebAssembly via Pyodide, pushing inference into the visitor's browser with no server, bounded by CPU compute and browser memory.
Cloudflare AI Gateway spend limits cap dollar spend per rule and return HTTP 429 when exceeded, turning cost control into an inline circuit breaker, not a billing alert.
Vercel's CDN now serves a cached copy when an origin fails, but only inside the stale-if-error window you set. It masks transient failures, it does not fix the origin.
Vercel's cache-tag purge API clears CDN, Runtime, and Data cache at runtime, no redeploy. The verb sets origin load: invalidate stays safe, dangerously-delete can stampede.