The Gap Between Ship and Shipped

The Gap Between Ship and Shipped

When 80% of "shipped" features don't exist, builders build anywayKey Takeaways:Vercel announced 5 features, shipped 1 - that's an 80% vaporware rateAI Gateway works beautifully - proof that shippin...

AI & DevelopmentJuly 3, 20253 min readRamakrishnan Annaswamy

When 80% of "shipped" features don't exist, builders build anyway


Key Takeaways:

  • Vercel announced 5 features, shipped 1 - that's an 80% vaporware rate

  • AI Gateway works beautifully - proof that shipping right matters

  • Cloudflare and Invariant ship features, Vercel ships promises

  • $10 and 24 hours proved more about building than any announcement

  • Early access to builders drives innovation, gatekeeping kills it

  • The gap between promise and product is where builders live


A $10 experiment revealed the truth about modern platform launches.

Vercel Ship 2025 announced five features. One actually works.

  • AI Gateway - early product, available.

  • Queues - limited beta, broken community form link

  • Sandbox - limited beta

  • Micro-frontends - enterprise prison

  • BotID - invite only

This isn't about Vercel. It's about an industry that normalized promises but not production ready.


Repository: github.com/ramakay/vercelship25

Live demo: https://vercelship25.vercel.app/

const useAIGateway = {
  gateway: @vercel/ai-sdk,
  models: [gpt-4o, claude-3.5, gemini-1.5],
  features: [streaming, racing, judging],
  complexity: minimal
};

No API key juggling. No provider SDKs. Just model strings.

This is what builders need. Simple. Powerful. Available.

Ship implies it works. Ship implies deployment. Ship implies builders can use it today.


Meanwhile, the competition actually ships.

Cloudflare AI Gateway has caching that cuts costs 90%. Rate limiting that works. Load balancing. Failover. Today.

Invariant Labs has agent tracing. Production guardrails. Visual execution graphs. Today.

Not tomorrow. Not "coming soon." Today.

Platforms shipping features capture developers. Platforms shipping promises train competitors' customers.


24 hours. One working feature. What emerged? An AI battle arena.

Grok vs Claude vs Gemini competing in real-time. Streaming responses. Cost tracking to $0.0001. Automated judging.

// What worked
const workingFeatures = {
  aiGateway: true,
  modelRacing: true,
  streamingUI: true,
  costTracking: true
};

// What didn't
const missingFeatures = {
  queues: 'use localStorage',
  sandbox: 'build client-side',
  microfrontends: 'not available',
  botId: 'invite only'
};

Three models. One API. Actual competition.

No Queues? Used localStorage. No Sandbox? Built client-side. No proper async? Created animations that solve real UX problems.

Constraints don't limit builders. They reveal platform limitations.


The $10 budget forced architectural truth:

  • Stream everything to minimize tokens

  • Extract key insights for display efficiency

  • Build resilient for missing features

  • Degrade gracefully around gaps

Every workaround documented what platforms should build. Every animation replaced a missing feature. Every hack pushed the ecosystem forward.

The best critique of a platform is building on it anyway.


The repository serves three purposes:

  • Accountability - Working code gently pressures platforms toward shipping

  • Education - Documented workarounds prevent others from repeating discovery

  • Progress - Public solutions show what's needed through demonstration

Cloudflare didn't accidentally build better features. They responded to builders showing gaps through code.

Working code moves ecosystems forward faster than feature requests.


This connects to broader patterns in AI infrastructure.

Previous experiments showed AI models work better in teams. Gemini for architecture. O3 for logic. Claude coordinating. Consensus through orchestration beats single model responses.

This experiment reveals why that future stalls: the infrastructure isn't ready.

Vercel identified every pain point correctly. Unified model access matters. Async processing is essential. Sandboxing enables new paradigms.

They just haven't built the solutions yet.

Vision without availability is market education for competitors.


Modern software normalized incomplete releases. "Coming soon" became industry standard. Builders trained themselves to work around absence.

But progress happens when builders create despite constraints.

Every public workaround becomes a feature request written in code. Every creative solution documents a platform need. Every success despite limitations proves what's possible.

Beta now means "we thought about it." Public beta means "we have a landing page."


The repository contains:

  • Working AI Gateway patterns

  • Queue workaround strategies

  • Async UX solutions via animation

  • Token economics reality ($0.0001 precision)

  • Platform limitation documentation

Fork it. Break it. Improve it. Show what's needed by building it.


For platforms: Beta should mean builders can test. Limited should mean documented limits. Ship should mean available now. Early access drives innovation, not gatekeeping.

For builders: Create with what exists. Share every workaround. Vote with working code. Push platforms through building.

The frontier isn't where platforms announce. It's where builders make it.


Ship should mean shipped.

Early access should mean access.

Until then, builders build anyway.

And platforms that enable them win.


The future belongs to platforms that empower builders with early access, not those that gate innovation behind invitations.

The code is the call to action.

RA

Ramakrishnan Annaswamy

Principal Architect

Platform DevelopmentAI InfrastructureDevelopment ToolsPlatform Strategy