Near-Term Predictions: 2026–2028

Learning Objectives

• Identify the near-term AI trends most supported by current evidence
• Explain why model cost collapse matters for AI adoption across the economy
• Distinguish between high-confidence near-term predictions and more speculative medium-term trajectories

Calibrating Confidence

Before making predictions about AI's future, it is worth calibrating what "near-term" means and how confident we should be.

High confidence: Trends already clearly underway, with strong momentum from multiple independent directions — technology, economics, regulatory, and behavioral.

Medium confidence: Trends with clear directional evidence but meaningful uncertainty about pace and scale.

Low confidence / speculative: Possibilities that could occur but depend on breakthroughs or conditions that may or may not materialize.

This section focuses on high-confidence near-term predictions — things we can see clearly from where we stand in early 2026. The further out you predict in AI, the lower your confidence should be.

Agentic AI Becomes Standard Knowledge Worker Tooling

The clearest near-term trend is the mainstreaming of agentic AI — systems that don't just respond to questions but autonomously execute multi-step tasks.

In 2023–2024, AI assistance primarily meant: ask a question, get an answer. In 2025–2028, AI assistance increasingly means: describe a goal, come back when it's done.

What this looks like in practice:

• An AI agent that reads your email, schedules meetings, drafts responses, and flags items requiring your attention — running continuously in the background
• A coding agent that receives a GitHub issue, implements a fix, writes tests, and opens a pull request — with a human reviewing the result
• A research agent that synthesizes information from dozens of sources into a briefing document you review before a meeting
• An AI sales development representative that handles initial outreach, qualifies leads, and hands off to human account executives

Salesforce's Agentforce has already deployed AI SDRs at enterprise scale, reaching 12,000 customers and $500 million ARR by early 2026. OpenAI's Operator handles web-based tasks autonomously. Anthropic's Claude Code — with Voice Mode, Computer Use, Agent SDK, and 1 million token context — represents the coding agent paradigm becoming accessible to non-experts. GitHub Copilot's Coding Agent now autonomously implements issues and opens pull requests. Frameworks like LangGraph v1.0, CrewAI (processing 12 million+ agent executions daily), and AG2 (the community fork of AutoGen) provide the infrastructure layer for building custom agents.

Confidence level: High. This is not a prediction — it is already happening. The question is pace of adoption, not direction.

Multimodal Becomes the Universal Default

In 2026, users already expect AI systems to handle text, images, and increasingly audio and video. By 2028, the expectation will be that AI handles all modalities natively — as a baseline, not a premium feature.

What this means:

• You photograph a document and ask AI to summarize it — standard capability
• You upload a video of a machine problem and ask for diagnosis — routine
• You speak naturally to an AI assistant that understands context, emotion, and interruptions — normal
• AI natively analyzes charts, graphs, and visualizations without requiring text transcription first

Current state: GPT-5.5, Gemini 3.1 Pro, Claude Opus 4.7 (with 1 million token context), and Grok 4.1 already have strong multimodal capabilities. The 2026–2028 transition is about these capabilities becoming the expected baseline rather than a differentiating feature.

Confidence level: High. The technology exists; the question is UX refinement and deployment scale.

AI Embedded in Every Significant Software Product

By 2028, any software product that does not have AI capabilities will feel dated — similar to how a software product today that doesn't have a mobile app feels dated.

Areas of rapid embedding:

• Productivity suites: Microsoft 365 Copilot, Google Workspace AI — AI writing, summarizing, analyzing in Word, Excel, Gmail, Docs
• CRMs: Salesforce, HubSpot — AI drafting emails, summarizing deals, predicting churn
• Developer tools: GitHub Copilot (Coding Agent, CLI, Mission Control), Cursor ($2 billion+ ARR, $29.3 billion valuation), AWS Kiro (spec-driven development), JetBrains AI — code completion, agentic development, and review as standard
• Design tools: Figma AI, Adobe Creative Cloud AI, Canva AI — generating assets, editing, style transfer
• Communications: Zoom AI, Slack AI, Teams Copilot — meeting summaries, action items, search across conversations

The scale of investment behind this embedding is staggering: the Stargate Project alone represents $400 billion+ in AI infrastructure across five US sites, with the Abilene, Texas facility already operational. Custom AI chips — OpenAI's Titan, Amazon's Trainium3 (3nm), and Google's TPU Ironwood v7 — are being designed specifically to power this next generation of embedded AI.

The strategic implication: AI fluency becomes a prerequisite for using the tools professionals use every day — not a separate AI skill, but embedded in the tools themselves.

Confidence level: High. This is already happening across enterprise software.

Vertical Integration: The Terafab Wave

The 2026–2028 window is also when the largest AI buyers start to own their own silicon supply — not just design chips, but own the foundry. The flagship example is Terafab, the $20-25 billion joint venture announced March 21, 2026 by Tesla, SpaceX, and xAI, with Intel joining as the process-technology partner on April 7, 2026. Terafab targets one terawatt of AI compute output per year — roughly 50× current global AI chip production — on Intel's 18A and 2 nm nodes.

Terafab will produce Tesla's AI5 and AI6 chips (taped out April 2026 for FSD and Optimus), restart Tesla's Dojo3 program now repositioned for space-based AI compute, xAI training accelerators for the post-Colossus 2 era, and SpaceX's radiation-tolerant orbital inference silicon. The pilot line targets 3,000 wafers per month by 2029, scaling toward 1 million wafer starts per month long-term.

Why this matters: No single customer-owned foundry of this scale has been attempted. If Terafab delivers, it creates the first credible non-NVIDIA frontier training platform at scale. If it slips — and AI5 is already roughly two years late, while Dojo has pivoted twice in eight months — NVIDIA's moat holds and the custom-silicon thesis is weakened industry-wide.

Confidence level: Medium. The announcement and Intel partnership are confirmed, but Tesla's lack of foundry experience makes timelines uncertain. What is near-certain is that vertical-integration attempts from several frontier AI buyers (Terafab, Stargate's chip program, Amazon Trainium, Google TPU) will reshape the semiconductor industry over the 2026-2028 window.

The EU AI Act and International Governance Come Into Force

The EU AI Act — passed in 2024 — is being implemented in phases through 2026 and beyond. By 2027, most of its provisions will be in force across EU member states, affecting:

• Prohibited AI systems: Social scoring, subliminal manipulation, real-time biometric surveillance in public spaces
• High-risk AI systems: AI in hiring, lending, healthcare, and law enforcement face transparency, data governance, and human oversight requirements
• General-purpose AI models: Large models must document capabilities, risks, and copyright compliance

Beyond the EU, multiple other jurisdictions are developing AI governance frameworks — the UK's AI Safety Institute (renamed the AI Security Institute in 2025), Japan's $135 billion AI investment with a soft regulatory approach, South Korea's $735 billion AI commitment, and ongoing discussions at the United Nations about international AI safety standards. The EU AI Act timeline is now concrete: prohibited AI practices banned since February 2025, general-purpose AI model transparency requirements in effect from August 2025, and high-risk AI system requirements phasing in through 2026–2027.

Penalties are substantial: up to EUR 35 million or 7% of global turnover for prohibited practices. The bulk of remaining obligations — including high-risk AI system requirements and full enforcement — take effect in August 2026, with each EU member state required to establish at least one AI regulatory sandbox.

Confidence level: High. The regulatory direction is clear; specific implementation timelines may shift.

Model Cost Collapse: Frontier Intelligence Becomes 100x Cheaper

One of the most profound near-term changes is happening in AI economics.

The cost to run a frontier-class AI query has been falling approximately 10x per year since 2020. DeepSeek's efficiency breakthrough in early 2025 accelerated this trend. By 2027–2028:

• What costs $100 per million tokens today will cost $1 or less
• Tasks currently too expensive to automate (reviewing every customer email, analyzing every transaction in real time) will become economically viable
• AI adoption barriers will shift from cost to organizational change management

Historical context: GPT-3 API access in 2021 cost approximately $60 per million tokens. By early 2026, frontier-class intelligence costs $3–15 per million tokens — and pricing drops have been dramatic: Claude Opus 4.7 dropped from $15/$75 to $5/$25 per million tokens, while DeepSeek's efficiency breakthroughs pushed costs even lower for open-weight models. The trajectory continues downward.

Confidence level: High. Hardware efficiency, software optimization, and competition all point in the same direction.

Physical AI: Robots and Autonomous Vehicles at Scale

Autonomous vehicles are not a future technology — they are a present technology in specific contexts. By 2028:

• Waymo is operating fully autonomous robotaxis across 10 US cities, having completed 170 million+ autonomous miles with no safety driver
• Zoox (Amazon) launched its commercial robotaxi service in San Francisco and Las Vegas, with expansion to Austin and Miami announced in March 2026, and a partnership with Uber to make Zoox robotaxis available on the Uber network
• Aurora Innovation has commercially deployed driverless trucking across 10 routes connecting Dallas, Houston, Phoenix, El Paso, and more — 250,000+ driverless miles with zero Aurora Driver-attributed collisions, targeting 200+ trucks by end of 2026
• Tesla FSD continues expanding, with the dedicated Cybercab entering production in April 2026
• Robotaxis are operating in 20+ cities globally, with multiple operators scaling rapidly

Humanoid robots are at an earlier but real commercial stage:

• Figure AI (valued at $39 billion) has progressed to Figure 03, with humanoid robots deployed in factory settings with BMW and others
• Tesla Optimus Gen 3 has over 1,000 units deployed across Tesla factories, with mass production started January 2026, 50 actuators per hand (25 per hand), and a target production cost of $20,000 per unit — external customer deliveries anticipated late 2026
• Boston Dynamics Atlas (electric) is now commercially deployed in warehouse and logistics tasks
• The first commercial-scale humanoid deployments are underway in controlled industrial environments, with multiple companies scaling production

What is NOT happening yet: Humanoid robots in homes, general-purpose manipulation in unstructured environments, or humanoid robots meaningfully displacing manufacturing employment at scale. These are medium-term possibilities, not near-term certainties.

Confidence level: Medium-high. Autonomous vehicles are certain to continue expanding; humanoid robot timelines are harder to predict.

Orbital AI Data Centers Emerge as a Real Category

What was science fiction in 2024 moved into FCC filings — and onto an IPO roadshow — by 2026. On January 31, 2026, SpaceX filed for authorization to deploy up to one million satellites functioning as orbital AI data centers, with a long-term target of 100 GW per year of space-based power generation. Each Starlink V3 satellite — launching 60 per Starship starting 2026, with 1 Tbps download capacity and terabit-class laser mesh networking — is designed to serve as a distributed orbital compute node.

The bet became central to SpaceX's identity over the course of 2026. In February, SpaceX absorbed Elon Musk's AI lab xAI (maker of the Grok models) in a roughly $250 billion all-stock deal, folding frontier-model software into the same company that builds the rockets and satellites. By June, as SpaceX headed toward one of the largest IPOs in history — valuing the combined company at roughly $1.5 trillion — Musk made orbital compute the centerpiece of his pitch. At a JPMorgan investor roadshow, he argued that solar-powered data centers in space could become "the primary means by which AI can be expanded," sidestepping the power-generation and permitting bottlenecks that increasingly constrain data centers on the ground. SpaceX told investors it plans to spend about $12.7 billion on AI in 2026 alone.

The Musk ecosystem's silicon roadmap reflects the bet: Tesla's restarted Dojo3 program was explicitly repositioned in January 2026 for "space-based AI compute," and Terafab will produce the radiation-tolerant inference silicon needed for orbital deployment.

The compute business is already real on the ground. Even before a single orbital data center comes online, SpaceX has become a major compute landlord. In June 2026 it agreed to rent roughly 110,000 Nvidia chips to Google for about $920 million a month — close to $30 billion over three years — as "bridge capacity" while Google's own data centers scale, on top of an even larger compute arrangement with Anthropic. The terrestrial compute-leasing revenue is what is meant to fund the orbital ambition.

What this actually enables (near-term): distributed edge inference for remote/mobile/underserved regions, reduced terrestrial data-center energy draw, and a physically diversified compute layer for defense and disaster-resilience scenarios. What it does NOT enable yet: frontier model training in orbit (the power, thermal, and data-throughput requirements are far beyond current satellite capabilities).

Confidence level: Medium. The FCC filing, satellite launches, silicon repositioning, and multi-billion-dollar compute deals are all confirmed, but the economics of orbital compute versus terrestrial data centers remain unproven. Analysts who watched the IPO roadshow noted that Musk's space-data-center pitch was long on vision and short on operational detail. Skeptics add that ground-based solar plus batteries plus Stargate-scale buildouts may undercut orbital compute on cost per FLOP for most workloads through the end of the decade.

Real-Time AI Glasses Go Mainstream

Smart glasses with AI assistants are transitioning from novelty to mainstream:

• Ray-Ban Meta smart glasses have sold 7 million+ units (2025), with Llama 4 AI integration, live translation, and image recognition. The premium Ray-Ban Meta Display ($799) adds a built-in teleprompter and Meta Neural Band (EMG wristband) for gesture control — demand so high that international expansion was paused
• Samsung Galaxy XR headset launched October 2025 ($1,800), the first device on Android XR (Google's new platform) with Gemini AI built in. Samsung and Google are developing lightweight AR smart glasses for 2026 with audio prompts, notifications, and live translation
• Apple Vision Pro has struggled commercially (~390,000 units shipped in 2024, production paused in early 2025), but Apple is reportedly redirecting effort toward a lighter smart glasses form factor
• The ability to look at something and ask AI a question — without looking at a phone — is becoming a natural interaction, with at least five Android XR devices expected in 2026

Confidence level: Medium-high. Ray-Ban Meta's commercial success validates the form factor; the ecosystem is expanding rapidly even as Apple's premium headset approach struggles.

Key Takeaways

• Agentic AI (goal-directed, autonomous) is already mainstream in coding and research; it will be table-stakes for knowledge work by 2028
• Multimodal AI becomes the universal baseline — text, images, audio, video handled natively by all major models
• Vertical-integration chip plays like Terafab ($20-25 billion Tesla/SpaceX/xAI/Intel foundry) reshape the semiconductor industry — if they execute
• Orbital AI data centers move from sci-fi to FCC filings — Starlink V3 and Dojo3's space-based compute pivot mark the start of a new deployment category
• Model cost collapse continues: frontier-class AI will be 10-100x cheaper by 2028, making new applications economically viable
• The EU AI Act creates binding regulatory requirements affecting high-risk AI globally; international governance frameworks are emerging
• Physical AI (autonomous vehicles, humanoid robots) is arriving in specific industrial contexts — not yet general-purpose