Developer Tools & Platforms — 2026年6月14日 週次レポート
重要な発見
重要な発見（11件）
- 1.DeepSeek V4 captured 17% of Vercel AI Gateway token share in a single month following its launch, rising from under 1%, while its spend share remained near 1% — reflecting a price of $0.14 input / $0.28 output per million tokens, roughly 20–50× lower than comparable Anthropic models. [1]
- 2.Anthropic's share of Vercel AI Gateway spend grew from 61% to 65% in May 2026, holding 70–80% of spend across every high-stakes use case including AI app generation, back-office agents, and coding agents, even as token volume fragmented. [1]
- 3.Total AI Gateway tokens grew +20% MoM and total spend grew +43% MoM in May 2026, with customers paying an average of almost 20% more per token than in April — indicating spend concentration toward higher-quality frontier models. [1]
- 4.Vercel's agentic infrastructure platform accumulated a third major production customer case study: Okara, a 4-person team, processes 4 billion tokens daily on Vercel, managing AI CMOs for 120,000+ businesses. [2]
- 5.Teams routing at 1M+ requests use 11 or more distinct models, while teams at 10M+ requests average 35 distinct models in regular use — reflecting systematic multi-model routing as operational cost discipline. [4]
- 6.Gemini 3.5 Flash launched at a higher price than Gemini 3.0 Flash in May 2026, with migration failing to materialize at scale: by month-end, Gemini 3.5 held only 7% of the Flash family's tokens while Gemini 3.0 held 90% — a clear signal that price is a decisive factor in production model adoption. [1]
- 7.Vercel documented an inference theft attack on April 12, 2026 in which traffic spiked to 1,300 requests per minute at ~10x normal volume, representing an inference cost run rate of over $10,000 per day, delivered through residential proxies that defeated standard per-IP rate limits. [5]
- 8.Supabase's $500M Series F at a $10B pre-money valuation, led by GIC and announced June 4, 2026, continues to shape the open-source backend infrastructure market; new product releases this period include Multigres v0.1 Alpha and npm supply chain attack protections. [6]
- 9.AWS's redesigned Amazon Bedrock console (June 4, 2026) supports the OpenAI Responses API, OpenAI Chat Completions API, and Anthropic Messages API, enabling existing OpenAI- or Anthropic-based applications to run on Bedrock with minimal code changes. [7]
- 10.Atlassian's Teamwork Lab research finds that workers who disclose AI use are judged 10x lazier than peers doing identical work unless their company has built a culture that celebrates it — a behavioral finding with direct implications for enterprise AI adoption strategies. [9]
- 11.Cloudflare's Security Insights system now processes over 120 scans per second, a 10x increase in global scanning capacity achieved by optimizing Kafka consumers, Postgres queries, and its API without adding hardware, as reported June 12, 2026. [10]
エグゼクティブサマリー（8件）
- •AI model spend fragmentation is the defining new development of this period: DeepSeek V4 captured 17% of Vercel AI Gateway token share in a single month, rising from near zero, while Anthropic simultaneously grew its spend share from 61% to 65% — revealing a two-tier market where low-cost models absorb volume and frontier models capture value. [1]
- •Total AI Gateway spend grew +43% MoM in May 2026 against +20% token growth, with customers paying ~20% more per token on average than in April — indicating that spend is accelerating faster than usage as teams shift to higher-quality, more expensive models for production workloads. [1]
- •Agentic infrastructure has accumulated a third major production validation point: Okara, a team of four people, processes 4 billion tokens daily on Vercel while managing AI CMOs for 120,000+ businesses — demonstrating that agentic infrastructure now enables extreme operational leverage at small team sizes. [2]
- •Cost-routing discipline is emerging as a distinct engineering practice: teams deliberately route high-volume, lower-risk work to cheaper models while reserving frontier models for quality-critical tasks, with multi-model usage scaling to 11+ models at 1M requests and 35+ models at 10M+ requests. [1] [4]
- •Supabase's $500M Series F at a $10B valuation remains a defining competitive event, now compounded by continued product shipping including Multigres v0.1 Alpha and npm supply chain protections, deepening its position as the default database layer for agentic applications. [6]
- •Atlassian deepened its AI-native toolchain this period by enabling direct assignment of Jira work items to Cursor as an agent orchestration platform and adding Claude Code support to Bitbucket Agentic Pipelines — further collapsing the separation between project planning and code execution layers. [9]
- •Inference theft remains an active and materially significant AI security threat: a single prompt to a frontier model can cost $2, making AI inference approximately one million times more expensive per call than standard HTTP requests, and residential proxy-based attacks can bypass standard rate-limiting defenses. [5]
- •AWS's Bedrock API compatibility layer — supporting OpenAI and Anthropic client libraries with minimal code changes — continues to represent the most structurally significant competitive move by a hyperscaler to capture the installed base of developers already invested in OpenAI or Anthropic tooling. [7]
市場動向
Agentic Infrastructure: Production Scale Deepens with New Customer Evidence
The agentic infrastructure trend established last period continues to accelerate with new production-scale evidence. Vercel reported that weekly deployments doubled in three months, with over 30% of deployments initiated by coding agents — up 1000% from six months prior — and Claude Code accounting for 75% of agent-driven deployments. Projects deployed by coding agents are 20 times more likely to call AI inference providers than those deployed by humans. [3] New this period, Okara — a 4-person t…
AI Model Spend Fragmentation Intensifies: DeepSeek V4 Captures 17% Token Share in One Month
Vercel's AI Gateway production index for May 2026 reveals a significant new development in the AI provider landscape. DeepSeek's share of tokens jumped from under 1% to 17% in a single month following the launch of DeepSeek V4, while its share of spend stayed near 1%. DeepSeek V4 Flash launched at $0.14 input / $0.28 output per million tokens, roughly 20–50× lower than comparable Anthropic models. [1] Anthropic's share of spend grew from 61% to 65% in May, holding 70–80% of spend across every hi…
Cost Discipline Becomes a Routing Strategy as AI Spend Grows
According to Vercel's May 2026 AI Gateway production index, teams are increasingly implementing smarter model routing strategies rather than simply increasing budgets. Teams sent high-volume, lower-risk work to cheaper models while reserving frontier models for quality-critical tasks. A clear example: Gemini 3.5 Flash launched in May at a higher price than Gemini 3.0 Flash, but migration didn't happen at scale — by month-end, 3.5 held only 7% of the Flash family's tokens while 3.0 held 90%. This…
Inference Theft Remains an Active and Evolving AI Security Threat
Vercel documented a real attack on April 12, 2026 in which traffic to its docs AI chat endpoint spiked to roughly ten times normal volume, reaching 1,300 requests per minute at peak — representing an inference cost run rate of over ten thousand dollars per day — delivered through residential proxies that defeated standard per-IP rate limits. Vercel noted that a single prompt to a frontier model can cost $2, making AI inference approximately one million times more expensive per call than standard…
Supabase $500M Series F Cements Open-Source Backend Infrastructure as a Standalone Market
Supabase announced a $500M Series F at a $10B pre-money valuation, led by GIC, as of June 4, 2026. [6] (company announcement — may reflect promotional framing) This follows a sustained period of product expansion including becoming an official ChatGPT app (May 8, 2026), achieving ISO 27001 certification, and launching Multigres v0.1 Alpha, described as an operating system for Postgres. The scale of the funding round signals strong investor conviction in open-source backend infrastructure as a fo…
AWS Bedrock API Compatibility Layer Lowers Migration Barrier from OpenAI and Anthropic
AWS launched a redesigned Amazon Bedrock console on June 4, 2026, optimized for the bedrock-mantle endpoint which supports the OpenAI Responses API, OpenAI Chat Completions API, and the Anthropic Messages API — enabling customers to run existing OpenAI- or Anthropic-based applications on Amazon Bedrock with minimal code changes. [7] On June 1, 2026, AWS also added Amazon CloudWatch metrics for the bedrock-mantle endpoint, published under the AWS/BedrockMantle namespace, covering inference counts…
Atlassian Deepens AI-Native Toolchain with Cursor Integration and Agentic Pipelines
Atlassian announced the introduction of Cursor in Jira, enabling work items to be assigned directly to Cursor from Jira as an agent orchestration platform. Bitbucket Agentic Pipelines now supports Claude Code, extending the prior launch of Agentic Pipelines for automating repetitive engineering chores. Atlassian also reported that AI Alert Grouping in Jira Service Management saved 839 hours in 28 days. [9] New research from Atlassian's Teamwork Lab finds that workers who disclose using AI are ju…
Cloudflare Scales Security Infrastructure 10x Without Additional Hardware
Cloudflare's Security Insights system now processes over 120 scans per second, achieving a 10x increase in global scanning capacity by optimizing Kafka consumers, Postgres queries, and its API — without adding hardware. This was reported on June 12, 2026. [10] This development signals that infrastructure-layer security platforms are investing in throughput efficiency as the volume of production AI workloads and associated security scanning requirements grows, a trend relevant to developer platfo…
競合動向
Vercel: New Production Customer Data Extends Agentic Infrastructure Validation
Vercel's agentic infrastructure positioning continues to accumulate production-scale evidence. New this period, Okara — a 4-person team — processes 4 billion tokens daily across a multi-provider AI stack on Vercel, managing AI CMOs for 120,000+ businesses, with new AI models made available to users the same day they ship via AI Gateway. [2] Vercel's June 2026 AI Gateway production index shows total tokens grew +20% MoM and total spend grew +43% MoM in May, with DeepSeek V4 capturing 17% of token…
Supabase: Series F Funding Consolidates AI Ecosystem Integration Strategy
Supabase raised a $500M Series F at a $10B pre-money valuation led by GIC, announced June 4, 2026. [6] (company announcement — may reflect promotional framing) This follows the prior period's AI assistant integrations (official Claude connector in February 2026, official ChatGPT app in May 2026) and enterprise compliance milestones (ISO 27001 certification). New product releases this period include Multigres v0.1 Alpha, described as an operating system for Postgres, and protections against npm s…
Atlassian: Cursor in Jira and Claude Code in Bitbucket Extend Third-Party Agent Integrations
Atlassian announced the introduction of Cursor in Jira, enabling work items to be assigned directly to Cursor from Jira as an agent orchestration platform, and Bitbucket Agentic Pipelines now supports Claude Code. [9] Atlassian also reported that Rovo Dev Standard customers can use Claude Opus 4.7 with a lower credit multiplier for a limited time, and that Loom resolved over 80% of support inquiries with AI and reduced the likelihood of churn by 11% using Atlassian Customer Service Management. […
ソース活動
重要な変化の整理
DeepSeek V4 Captures 17% of AI Gateway Token Share in One Month
新規DeepSeek's share of tokens on Vercel AI Gateway jumped from under 1% to 17% in May 2026 following the launch of DeepSeek V4, while its spend share stayed near 1%. DeepSeek V4 Flash launched at $0.14 input / $0.28 output per million tokens, roughly 20–50× lower than comparable Anthropic models. [1] This is a new development not present in the prior period's April data and represents the first time a low-cost model has cleared the quality bar for production workloads at this scale on the gateway.
Vercel Agentic Infrastructure Validated by Additional Production Customer Data
更新Vercel's agentic infrastructure positioning, first established in April 2026 and validated last period by Superset and General Intelligence metrics, has been further extended this period by Okara: a 4-person team processing 4 billion tokens daily and managing AI CMOs for 120,000+ businesses on Vercel. [2] Total AI Gateway tokens grew +20% MoM and spend grew +43% MoM in May 2026. [1] The trend has evolved from two customer case studies to a broader pattern across multiple production deployments.
Supabase $500M Series F Continues to Shape Backend Infrastructure Market
継続監視Supabase's $500M Series F at a $10B pre-money valuation, led by GIC and announced June 4, 2026, remains a defining market development. [6] No new funding or valuation updates were found in this period's sources. The company continues to ship product updates including Multigres v0.1 Alpha and npm supply chain attack protections, consistent with the prior period's trajectory.
AWS Bedrock API Compatibility Layer Remains Active Competitive Move
継続監視AWS's redesigned Amazon Bedrock console (June 4, 2026) and CloudWatch metrics for the bedrock-mantle endpoint (June 1, 2026), supporting OpenAI and Anthropic-compatible APIs, continue to represent an active competitive strategy to capture developers already invested in OpenAI or Anthropic tooling. [7] [8] No new Bedrock API compatibility announcements were found in this period beyond what was documented last period.
Atlassian AI-Native Toolchain Expansion Continues with Cursor and Claude Code
継続監視Atlassian's sustained AI-native platform expansion continues with Cursor in Jira for direct agent work assignment and Bitbucket Agentic Pipelines support for Claude Code. [9] These developments are consistent with the prior period's trajectory and do not represent a significant new strategic direction, confirming this as a continuing trend rather than a new escalation.
示唆・見るべき論点（9件）
- 1.The divergence between DeepSeek V4's 17% token share and ~1% spend share on Vercel AI Gateway establishes a structural 'token-spend gap' dynamic: low-cost capable models will absorb the bulk of volume in production environments while frontier models capture the economic value. Developer platforms should architect for this bifurcation by providing model-routing primitives that make tier-selection explicit and measurable, not just cost-efficient. [1]
- 2.The failure of Gemini 3.5 Flash to displace Gemini 3.0 Flash despite being a newer model — holding only 7% of Flash family tokens at month-end — confirms that production teams exhibit strong price inertia when upgrading within a model family unless quality improvements are demonstrably compelling. This has strategic implications for AI model pricing strategies: launching a successor at a higher price point risks permanent volume fragmentation within a vendor's own product line. [1]
- 3.Okara's case — a 4-person team processing 4 billion daily tokens and serving 120,000+ businesses — represents a new benchmark for agentic operational leverage. This pattern suggests that team size is becoming decoupled from operational scale in AI-native companies, with infrastructure abstraction (multi-provider routing, zero-downtime model switching) serving as the key enabler. Platforms that can demonstrate this leverage in sales cycles will have a structural advantage over those offering only…
- 4.Multi-model routing is transitioning from an optimization tactic to a core architectural discipline: at 10M+ requests, teams already use an average of 35 distinct models. This creates durable demand for AI gateway infrastructure as a standalone product category — not a feature — and signals that observability, cost attribution, and performance benchmarking across heterogeneous model portfolios will become critical enterprise requirements. [4]
- 5.Atlassian's finding that AI-disclosing workers are perceived as 10x lazier unless company culture actively celebrates AI use is a significant enterprise adoption risk factor. Developer platform vendors targeting enterprise sales should consider embedding AI adoption culture guidance into their go-to-market and customer success motions, not just technical enablement, as cultural resistance may slow deployment velocity independently of technical readiness. [9]
- 6.Cloudflare's 10x scanning capacity increase without additional hardware — via software optimization of Kafka consumers and Postgres queries — signals that infrastructure-layer security platforms are entering a phase of efficiency-driven scaling. As agentic workload volumes grow, the ability to scale security scanning throughput without proportional infrastructure cost becomes a durable competitive differentiator for platforms serving large-scale deployments. [10]
- 7.The inference theft threat documented by Vercel — where a single frontier model call costs $2 versus $2/million for standard HTTP requests — represents an asymmetric economic attack surface with no direct analog in prior developer security categories. As AI endpoints proliferate across developer platforms, per-request verification infrastructure will become a non-negotiable security requirement, creating a distinct product category opportunity for security vendors specializing in AI workload pro…
- 8.Supabase's combination of $500M in fresh capital, official ChatGPT app status, ISO 27001 certification, and continued product shipping (Multigres v0.1 Alpha) creates a compounding moat in open-source backend infrastructure. Competitors lacking comparable AI ecosystem distribution channels and enterprise compliance credentials will find it increasingly difficult to compete for agentic application workloads at the database layer. [6]
- 9.AWS's zero-refactoring migration path for OpenAI and Anthropic client library users on Bedrock is a systematic attack on developer inertia as a competitive moat. By eliminating the primary switching friction — the cost of re-implementing API clients — AWS positions Bedrock as a risk-free enterprise alternative, accelerating pressure on model providers to compete on model quality and safety rather than tooling lock-in. [7]
信頼度サマリー
今週追跡された 10 件のソース15 件の監視対象 URL から、期間中に新着・更新が検出された記事数。
各ソースは信頼度レベルに応じて重み付けされています。単独ソースの主張は AI 合成時に未検証としてフラグ付けされます。
ソース
Vercel's May 2026 AI Gateway production index reveals DeepSeek V4 captured 17% of token share in a single month, Anthropic's spend share grew from 61% to 65%, total tokens grew +20% MoM, total spend grew +43% MoM, and customers paid ~20% more per token on average than in April.
関連: Market Trends / Competitor TrendsOkara, a 4-person team, processes 4 billion tokens daily across a multi-provider AI stack on Vercel, managing AI CMOs for 120,000+ businesses and accessing new AI models the same day they ship via AI Gateway.
関連: Market Trends / Competitor TrendsVercel reported weekly deployments doubled in three months, with over 30% initiated by coding agents — up 1000% from six months prior — and Claude Code accounting for 75% of agent-driven deployments. Projects deployed by coding agents are 20x more likely to call AI inference providers.
関連: Market TrendsVercel's AI Gateway production index from 200K+ unique teams shows teams at 10M+ requests average 35 distinct models in regular use, and at 1M+ requests the majority route across 11 or more models.
関連: Market Trends / Competitor TrendsVercel documented an inference theft attack on April 12, 2026, spiking to 1,300 requests/minute via residential proxies, with an inference cost run rate of over $10,000/day. A single frontier model prompt can cost $2, making AI inference ~1 million times more expensive per call than standard HTTP requests.
関連: Market TrendsSupabase announced a $500M Series F at a $10B pre-money valuation led by GIC, with new product releases including Multigres v0.1 Alpha and npm supply chain attack protections. Company announcement — may reflect promotional framing.
関連: Market Trends / Competitor TrendsAWS launched a redesigned Amazon Bedrock console on June 4, 2026, supporting the OpenAI Responses API, OpenAI Chat Completions API, and Anthropic Messages API via the bedrock-mantle endpoint for minimal-code-change migration.
関連: Market Trends / Competitor TrendsAWS added CloudWatch metrics for the bedrock-mantle endpoint on June 1, 2026, covering inference counts, input/output token totals, and client error counts at account, project, model, and project-and-model granularity.
関連: Market TrendsAtlassian announced Cursor in Jira for direct agent work assignment, Claude Code support in Bitbucket Agentic Pipelines, AI Alert Grouping saving 839 hours in 28 days, and Teamwork Lab research finding AI-disclosing workers are judged 10x lazier unless company culture celebrates AI use.
関連: Market Trends / Competitor TrendsCloudflare's Security Insights system achieved a 10x increase in global scanning capacity, processing over 120 scans per second by optimizing Kafka consumers, Postgres queries, and its API without adding hardware, reported June 12, 2026.
関連: Market Trends