🛠️ Tech Trends Brief — June 8, 2026
15 Technologies for Pacific Glazing Corp
📅 June 8, 2026
🔬 15 Technologies
⚡ MVP Experiments
🏢 PGC Focused
What it is
Anthropic's new flagship model — a hybrid reasoning system pushing the frontier for coding and AI agents, featuring a 1M context window. Released alongside Dynamic Workflows for Claude Code (adaptive execution paths) and an "effort control" knob. Released June 2, 2026. Anthropic now valued at $730-965B after confidential SEC filing.
What it's used for
Production coding, multi-step agent tasks, complex document analysis, and computer-use automation. The 1M context window means it can process entire project specs in one shot.
Trend tags
AnthropicClaude Opus 4.8frontier modelshybrid reasoningIPO1M context
MVP experiment
In Claude Code with Opus 4.8 selected: "Analyze this 200-page curtain wall spec and extract all glass type specifications, U-value requirements, and testing standards. Output as a structured table." Time: 30 minutes.
Why it matters for PGC
Opus 4.8 handles PGC's entire spec library in one context window. Hybrid reasoning means it's good at both creative problem-solving (installation sequencing) and precise calculation (glass loads, thermal performance). The IPO validates the company's staying power.
🟢 TRY NOW
What it is
Perplexity's "orchestrator" agent that splits AI work between your local device and the cloud. Launched on Mac, now on Windows (June 3). The hybrid inference system decides real-time what runs locally (simple tasks, private data) vs. what goes to frontier models in the cloud.
What it's used for
Desktop automation — files, apps, and the web — with intelligent cost/performance routing. Tasks needing privacy stay local; complex reasoning uses cloud models.
Trend tags
Perplexityhybrid inferencedesktop agentlocal AIorchestrator
MVP experiment
Install Perplexity Personal Computer on a PGC Windows machine. Task: "Find the latest GL-301 spec document in the shared drive, extract the glass thickness requirements, and email them to me." Time: 1 hour.
Why it matters for PGC
Hybrid local-cloud means sensitive project documents stay local while using frontier-model intelligence. This is the architecture that scales — private data stays on-prem, compute power comes from the cloud.
🟢 TRY NOW
What it is
YC-backed startup building "the AI workforce for construction design engineering." AI agents perform quality control on technical documents and drawings — learning your standards, applying building codes, and flagging issues before they reach the field.
What it's used for
Automated drawing review: curtain wall detail correct per IBC code? Does glazing spec match structural drawings? AI checks thousands of drawings in minutes instead of days.
Trend tags
construction AIYCdrawing QCbuilding codesdesign review
MVP experiment
Upload one PGC curtain wall shop drawing set to Structured AI (or LightTable, Fresco). Ask it to flag: code compliance issues, dimension inconsistencies, glass type mismatches. Time: 30 minutes.
Why it matters for PGC
Drawing QC is manual, expensive, error-prone. An AI layer catching code violations before fabrication saves rework costs. Structured AI is YC-funded — strong signal for quality.
🟢 TRY NOW
What it is
Cursor 3.7 introduces Design Mode — click, draw, or describe changes by voice to help agents update your UI. Combined with multi-model support (Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro), Cursor is now a model-agnostic AI IDE where frontier models compete to be the engine inside it. New Jira integration also landed. Released June 5, 2026.
What it's used for
AI-powered software development with natural UI manipulation. Voice-to-code, click-to-modify, multi-model agent workflows.
Trend tags
CursorAI IDEDesign Modevoice codingmulti-model
MVP experiment
In Cursor 3.7 with Design Mode: Open the PGC Field Tools repo. Say "Move the filter buttons to the right side of the page and make them purple." Watch it happen. Time: 15 minutes.
Why it matters for PGC
Design Mode means Steve or a PM can describe UI changes without writing code. The model-agnostic design means PGC picks the best model for each task.
🟢 TRY NOW
What it is
Google will pay SpaceX $920 million per month from October 2026 through June 2029 for access to ~110,000 NVIDIA GPUs at xAI data centers. The deal (~$30B total) was signed days before SpaceX's June 12 IPO. SpaceX had merged with xAI in February 2026.
What it's used for
Google gets guaranteed AI compute at mass scale. SpaceX monetizes GPU infrastructure ahead of IPO. Sets a pricing benchmark for enterprise AI compute.
Trend tags
SpaceXGoogleAI computeGPU infrastructureIPO
MVP experiment
At $920M/mo for 110K GPUs = ~$8,363/GPU/month. Compare to running one local RTX 4090 at ~$0.30/hr. Could PGC run 80% of AI tasks on one local GPU for $2,000/year? Time: 15 minutes.
Why it matters for PGC
This deal signals big AI compute is becoming an industrial commodity. The unit economics show cloud inference is expensive at scale. For PGC's workload (spec extraction, bid analysis), local inference is orders of magnitude cheaper.
🟡 WATCH — market signal; reinforces local inference thesis for PGC
What it is
High-throughput, memory-efficient LLM serving engine from UC Berkeley. Achieves 20x throughput vs. naive implementations. Features PagedAttention for near-zero memory waste, continuous batching, and tensor parallelism. The standard for serving open-source models in production.
What it's used for
Serving open-source LLMs (Llama, Qwen, Phi, Mistral) at production scale on modest hardware. Turns a single GPU into a multi-user model server.
Trend tags
vLLMLLM servingthroughputinferenceBerkeleyopen source
MVP experiment
pip install vllm && python -m vllm.entrypoints.openai.api_server --model phi-4-mini:3.8b. Send requests via curl, compare latency vs. Ollama. Time: 1 hour.
Why it matters for PGC
One machine running vLLM serves the whole office's AI needs. Combined with Ollama for prototyping and vLLM for production, PGC gets enterprise-grade local AI for zero monthly fees.
🟢 TRY NOW
What it is
New open-source model offering training costs at $1.25/hour — a dramatic reduction from the $10M+ it cost to train GPT-3. Combined with nanochat (Karpathy's full training stack in one Python file), custom model training has collapsed from millions to hundreds of dollars.
What it's used for
Fine-tuning and training custom AI models for specific business domains — without Silicon Valley budgets.
Trend tags
Orion-100Bcheap trainingopen sourcenanochatcustom models
MVP experiment
git clone https://github.com/karpathy/nanochat && cd nanochat. Prepare 100 PGC specifications as training data. python train.py --data ./pgc-specs.txt. Time: 2 hours to train a tiny model. See if it learns PGC-specific terminology.
Why it matters for PGC
Custom model training is now affordable. PGC could train a model that understands glazing-specific terminology, internal standards, and preferred suppliers — for the cost of a nice dinner.
🟢 TRY NOW
What it is
Six contech startups raised a combined $121M in a single week. Standouts: LightTable ($22M Series A) — AI preconstruction platform reading drawings to flag errors. AI field copilot — quality control, scoping, on-demand knowledge. August Robotics — robots for AI data center construction via DeWalt partnership.
What it's used for
AI-powered preconstruction, field automation, and construction robotics — all attracting serious venture capital.
Trend tags
contechfundingconstruction AILightTablepreconstruction
MVP experiment
Browse LightTable's product page. Upload one PGC project drawing. See what errors/omissions it catches that were missed in manual review. Compare to PGC's change order rate from drawing errors. Time: 30 minutes.
Why it matters for PGC
$121M in one week confirms construction AI is a real market. LightTable's preconstruction AI could catch costly glass dimension errors before fabrication.
🟢 TRY NOW — test LightTable or Fresco on a real PGC project this week
What it is
Cognition AI rebranded Windsurf as Devin Desktop (June 2) — shifting from AI code completion to a "command center for managing all your agents in one place." Powered by in-house SWE-1.5 model, introduces Agent Control Protocol (ACP) for multi-agent orchestration.
What it's used for
Multi-agent software development — one agent writes code, another reviews it, a third runs tests, all coordinated from one IDE.
Trend tags
Devin DesktopWindsurfmulti-agentIDEACP
MVP experiment
Install Devin Desktop. Create project: "Build a dashboard showing PGC's open change orders grouped by project, updated daily from a spreadsheet." Watch multi-agent system decompose, build, test. Time: 1 hour.
Why it matters for PGC
Multi-agent orchestration beyond single coding agents. Specialized agents for frontend, data processing, testing — all coordinated. Evaluate when PGC's dev workload justifies it.
🟡 WATCH — powerful but complex; evaluate when workload justifies multi-agent orchestration
What it is
Google DeepMind's Gemini 3.5 Pro, successor to Gemini 3.1 Pro. Now available inside Cursor and other AI IDEs. Enhanced reasoning, longer context, improved agentic capabilities. Google pushing users from open-source Gemini CLI to closed-source Antigravity CLI.
What it's used for
State-of-the-art AI reasoning, coding, multimodal analysis, agent orchestration. Competes with Claude Opus 4.8 and GPT-5.2.
Trend tags
Gemini 3.5Googlefrontier modelsDeepMindAntigravity
MVP experiment
In Cursor, switch to Gemini 3.5 Pro. Ask: "Review this curtain wall structural calculation for ASCE 7 wind load compliance." Compare reasoning quality vs. Claude Opus 4.8. Time: 30 minutes.
Why it matters for PGC
Model competition benefits PGC — Claude, GPT, and Gemini compete inside same tools, driving quality up and prices down. Test each on PGC-specific tasks.
🟡 WATCH — test when available; competition is good for quality and pricing
What it is
GitHub completed switch to usage-based token billing for Copilot (effective June 1). Developers pay per token consumed across inputs, outputs, and cached context. The old unlimited plan is gone. Signals that all AI tools will eventually be metered.
What it's used for
Monetizing AI code generation at scale. Sets pricing precedent for every AI developer tool. Companies calculating token budgets like cloud spend.
Trend tags
Copilottoken billingusage-baseddeveloper costsGitHub
MVP experiment
Request PGC's GitHub Copilot usage for May. Calculate June cost under token model. Compare: local alternatives (Codeium, Continue + Ollama, Cursor BYOK). Time: 20 minutes.
Why it matters for PGC
Every AI tool PGC evaluates will move to usage-based pricing. Understanding token economics now helps make better build-vs-buy decisions.
🟡 WATCH + CALCULATE
What it is
The vast collection of drop-in CLAUDE.md skills (55K+ ⭐). New additions include transitions.dev (12 CSS animations as skill file), andrej-karpathy-skills (4 coding principles), and domain-specific skills for data processing, security, design. With Claude Opus 4.8 Dynamic Workflows, skills now support adaptive execution paths.
What it's used for
Extending Claude Code without writing skills from scratch. The "app store" of the agent era — install domain expertise without training models.
Trend tags
Claude Codeskillsagent customizationopen source55K stars
MVP experiment
Install 3 skills and test: curl -o .claude/skills/design.md https://raw.githubusercontent.com/.../transitions.md. Ask Claude: "Make this PGC dashboard look professional." Time: 15 minutes.
Why it matters for PGC
The skills ecosystem makes Claude Code exponentially more capable for free. Design + data processing + code review skills combine without writing configuration.
🟢 TRY NOW
What it is
The SLM thesis continues: models under 7B parameters (Phi-4, Llama-3.2-3B, Qwen2.5-7B, Gemma-3) achieve GPT-3.5-to-GPT-4 performance on business tasks — running entirely locally. Gartner: 40% of enterprise workloads using industrial AI agents by end of 2026. Smaller specialized models consistently outperform general-purpose giants on specific domains.
What it's used for
Spec extraction, document classification, material lookup, RFI routing, data entry — 80% of a business's actual AI needs.
Trend tags
SLMsedge AIlocal inferenceon-devicegood enough
MVP experiment
ollama run phi4-mini:3.8b. Prompt: "Extract all glass types, thicknesses, U-values, and SHGC values from this spec." Compare output to GPT-4o. Track accuracy, speed, cost per query. Time: 30 minutes.
Why it matters for PGC
Phi-4 on a $1,000 laptop handles spec extraction, bid parsing, and material lookups — zero per-query cost, zero data leaving PGC's network.
🟢 TRY NOW
What it is
New open-source tool for sharing context across agents. Workgroups collaborate on projects by giving multiple agents access to a shared context store — one agent's output becomes another's input. Solves the "siloed agent" problem. (O'Reilly Radar, June 2026)
What it's used for
Multi-agent collaboration: Agent A does research → writes findings. Agent B reads findings → creates plan → writes plan. Agent C executes plan. All share common context.
Trend tags
Kanwasshared contextmulti-agentagent collaborationworkgroup
MVP experiment
Set up Kanwas. Agent A extracts specs from PDF. Agent B generates material list from those specs. Agent C formats output as a quote. Watch context flow between them. Time: 1 hour.
Why it matters for PGC
PGC's workflows are multi-step: spec → quantities → materials → labor. Kanwas shared context connects these steps into an automated pipeline.
🟢 TRY NOW
What it is
Wave of infrastructure investments: Coralogix ($200M) AI observability, ZeroDrift ($10M) AI reliability testing, Lowfat (91.8% token savings) agent cost optimization. O'Reilly Radar tracks "agent infrastructure" as central 2026 question. Cerebras IPO (May 14) and SpaceX roadshow (June 8) signal Wall Street appetite for AI infrastructure.
What it's used for
Plumbing for production AI agents: monitoring, cost control, reliability testing, observability. The "Cloudflare for AI" category emerging.
Trend tags
agent infrastructureobservabilitytoken optimizationCoralogixCerebras
MVP experiment
Map PGC's current software infrastructure. Where would an agent fit? Identify needed pieces (monitoring, logging, cost tracking) before deploying agents in production. Time: 30 minutes.
Why it matters for PGC
The companies building agent infrastructure are getting funded at massive scale — confirming the market. PGC can adopt agents now, scale with infrastructure later when tools mature.
🟡 WATCH — infrastructure maturing; adopt agents now, scale with infrastructure later
| # |
Technology |
What it is |
PGC Applicability |
Verdict |
| 1 | Claude Opus 4.8 | Anthropic hybrid reasoning, 1M context, Dynamic Workflows | 🟢 High — full spec library, IPO validates staying power | TRY NOW |
| 2 | Perplexity Personal Computer | Hybrid local-cloud desktop agent for Windows/Mac | 🟢 High — private data local, cloud brain for complex tasks | TRY NOW |
| 3 | Structured AI (YC) | AI drawing QC — building code compliance | 🟢 High — catch curtain wall errors before fabrication | TRY NOW |
| 4 | Cursor 3.7 Design Mode | AI IDE with voice, click, draw UI editing | 🟢 High — PMs can describe UI changes | TRY NOW |
| 5 | SpaceX $30B Google Deal | $920M/mo GPU compute benchmark | 🟡 Med — confirms local inference thesis | WATCH |
| 6 | vLLM | 20x throughput LLM serving on one GPU | 🟢 High — serve whole office with one machine | TRY NOW |
| 7 | Orion-100B | Open-source model training at $1.25/hr | 🟢 High — custom model for cost of dinner | TRY NOW |
| 8 | Contech $121M Week | LightTable, field copilots, funding boom | 🟢 High — test LightTable/Fresco on real drawings | TRY NOW |
| 9 | Devin Desktop | Windsurf rebrand to multi-agent IDE | 🟡 Med — powerful but complex | WATCH |
| 10 | Gemini 3.5 Pro | Google's next frontier model | 🟡 Med — test vs Claude for PGC tasks | WATCH |
| 11 | Copilot Usage Billing | GitHub per-token pricing from June 1 | 🟡 Med — calculate PGC impact | WATCH |
| 12 | Awesome Claude Skills | 55K+ ⭐ drop-in skills ecosystem | 🟢 High — extend Claude Code for free | TRY NOW |
| 13 | Small Language Models | Phi-4, Llama-3.2, etc. — 80% of tasks, zero cost | 🟢 High — daily AI needs met locally | TRY NOW |
| 14 | Kanwas | Cross-agent shared context store | 🟢 High — connects multi-step workflows | TRY NOW |
| 15 | Agent Infrastructure Boom | Coralogix, ZeroDrift, Lowfat — AI ops funding | 🟡 Med — infrastructure maturing for future scale | WATCH |