Many leading companies in the AI space have found that rising compute costs, though often cited as a barrier, are no longer the primary constraint to AI adoption. Instead, latency, flexibility, and capacity pose greater challenges. For example, food delivery company Wonder sees AI adding only a few cents per order but faces cloud capacity limits due to rapid growth. Wonder’s CTO, James Chen, explains that while large AI models are currently cost-effective, hyper-customized small models per individual remain too costly. The company encourages experimentation but finds budgeting AI costs unpredictable. Similarly, biotech company Recursion balances on-premises and cloud resources to optimize large-scale AI workloads. Recursion’s CTO, Ben Mabey, advocates for multi-year investments in compute infrastructure to foster innovation, noting on-premises solutions can be significantly more cost-efficient for large tasks. Overall, cost is shifting from an obstacle to managing the pace and scale of AI deployment and infrastructure sustainability.

Presented by Salesforce

Vibe coding, the emerging trend of using generative AI to create code from natural language prompts, has transformed rapid prototyping with its speed and creativity. However, experts like Mohith Shrivastava, principal developer advocate at Salesforce, caution that while vibe coding excels in fast, initial development, it often lacks the security, governance, and infrastructure needed for enterprise-grade applications. This gap can lead to vulnerabilities, technical debt, and challenges in maintaining and scaling enterprise software.

Vibe coding's main strength lies in accelerating early-stage innovation, especially in the user interface and experience layers (the "green zone"). But it presents risks when applied broadly, especially to critical business logic and data layers (the "red zone") where security and compliance are paramount. Shrivastava advocates using vibe coding selectively, supported by specialized enterprise AI tools that embed security and governance, and cautioning against unchecked AI use in core infrastructure.

Salesforce's Agentforce Vibes product exemplifies this approach by combining AI-assisted coding with enterprise controls. It enables accelerated, AI-augmented development across green and red zones while maintaining trust and compliance, helping major enterprises like Coinbase, CGI, and Globo build production-ready apps faster. This hybrid strategy blends innovative AI capabilities with disciplined engineering to achieve both agility and enterprise-grade stability.

At a recent VentureBeat AI Impact event, Val Bercovici, Chief AI Officer at WEKA, highlighted the growing capacity and cost challenges in AI deployment, particularly around latency and cloud dependency. He explained how AI is approaching a point similar to Uber's surge pricing, where real market rates for AI usage—especially inference—will soon replace current subsidized costs. This shift, possibly by 2027, will force the industry to prioritize efficiency.

Bercovici emphasized the exponential business value in increasing tokens but pointed out that costs and latency pose sustainability challenges, especially for high-stakes applications requiring accuracy and security. He described how AI agents operate in swarms, executing complex tasks with many prompt-responses, making latency a critical bottleneck that currently demands high subsidized pricing.

Reinforcement learning has emerged as a key paradigm, blending training and inference workflows to advance AI capabilities toward artificial general intelligence. On profitability, Bercovici noted there's no one-size-fits-all infrastructure approach; organizations must tailor their strategies—on-prem, cloud, or hybrid—based on evolving needs. Ultimately, understanding and optimizing unit economics at the transaction level will be vital. The future of AI lies in smarter, more efficient large-scale application rather than reduced usage.

Google recently removed its Gemma 3 AI model from the AI Studio platform following criticism from Senator Marsha Blackburn, who accused the model of fabricating defamatory falsehoods about her. Although Gemma remains accessible via API, Google decided to restrict AI Studio access because it was being used by non-developers asking factual questions, which was not the model's intended use. The situation underscores the dangers of relying on experimental developer models, especially as AI companies continue to face political scrutiny that affects their model deployments. Google’s Gemma models were designed for developers and researchers rather than consumers, and their removal from AI Studio raises important issues about project continuity and model control. Users of AI models should be aware that such experimental tools can be withdrawn at any time, potentially disrupting ongoing projects. This controversy also reflects broader challenges within the AI industry regarding model accuracy, governance, and the balance between innovation and risk management.

New York City-based startup Augmented Intelligence Inc (AUI), operating stealthily but gaining buzz, has raised $20 million in a bridge SAFE round at a $750 million valuation cap, bringing total funding close to $60 million. This round, completed quickly, precedes a larger funding round currently in advanced stages. AUI combines transformer-based language models with a new neuro-symbolic AI architecture that separates linguistic processing from symbolic task reasoning. This hybrid design offers deterministic task execution and policy enforcement, addressing limitations of conventional transformer-only conversational AI systems like ChatGPT. AUI's Apollo-1 model targets task-oriented dialogue for enterprise use cases, emphasizing reliability and regulatory compliance across sectors such as healthcare and finance. Apollo-1 is already in closed beta with Fortune 500 companies, supports deployment across cloud and hybrid environments, and allows integration via developer tools or APIs. The startup's approach enables faster onboarding and cost-efficient, deterministic AI implementations for mission-critical workflows.

For over thirty years, CPUs have relied on speculative execution to optimize processing pipelines by predicting the outcomes of operations and avoiding stalls. While revolutionary at the time, this method introduces inefficiencies such as wasted energy during mispredictions, increased complexity, and security vulnerabilities like Spectre and Meltdown. A new architectural approach replaces speculative guesswork with a deterministic, time-based execution model. This approach uses a simple cycle-accurate time counter to schedule instruction execution precisely when operands are ready and resources are available, eliminating the need for pipeline flushes and speculative rollbacks.

This deterministic framework is supported by recently issued patents and integrates naturally with RISC-V instruction sets, particularly enhancing matrix and vector computations. It offers scalable performance suitable for AI and high-performance computing workloads while maintaining lower costs and power consumption compared to traditional CPUs and specialized processors like Google's TPU.

Unlike speculative architectures that unpredictably reorder instructions and may waste cycles on mispredictions, deterministic CPUs ensure a predictable flow of instruction execution, improving efficiency and reliability. The design simplifies compiler scheduling, maintains compatibility with conventional programming models, and is well-suited to AI/ML workloads where consistent, high utilization of vector and matrix units is critical.

This emerging paradigm shift presents a promising path forward for CPU architecture, potentially superseding speculative execution as the standard by enabling energy-efficient, predictable, and high-throughput AI performance.

The emergence of AI signals a pivotal move away from an era fixated on data and computing power to what Canva's co-founder Cameron Adams calls the 'imagination era.' In this era, creativity powered by AI transforms ideas into actionable outcomes. Canva’s new Creative Operating System (COS) sets the stage, integrating AI deeply to provide a unified platform for content creation that spans documents, videos, social media, and more. Its AI model understands complex design elements and supports real-time collaboration, allowing seamless content generation and editing within one workflow. Innovations like 'Ask Canva' offer interactive design assistance through AI, enhancing creative collaboration. Canva’s growing ecosystem includes professional tools and marketing automation features that enable enterprises to maintain consistent branding and deploy campaigns efficiently. With over 250 million monthly users and major clients such as Walmart and Disney, Canva is positioned prominently amid competitors like Adobe Express and Figma, leveraging an extensive asset library and user-friendly interface. The company advocates a future where AI and humans collaborate, emphasizing openness and strategic use of proprietary and third-party AI models. This approach underlines the importance for IT leaders to recognize and adopt AI-powered creativity platforms to stay competitive.

Google Cloud has introduced Vertex AI Training, a managed Slurm environment designed to support enterprises in training their own AI models at scale. This new service offers access to a wide range of chips for large-scale model training, built-in data science tools, and automated job scheduling with automatic recovery for hardware failures. It targets organizations planning extensive training, including AI startups, enterprises customizing AI for internal processes, and sovereign groups developing region-specific models. Vertex AI Training provides flexibility for enterprises to build or customize models beyond fine-tuning, distinguishing itself by leveraging Google's expertise and offering robust monitoring and management capabilities. Existing competitors include CoreWeave, Lambda Labs, AWS, and Microsoft Azure, all providing GPU access for AI training workloads. Early adopters include AI Singapore and Salesforce's AI research teams, indicating strong interest in tailored AI solutions. The service aims to reduce the complexity and cost barriers of AI training, particularly for companies requiring large-scale compute resources with reliability and efficiency.

Anthropic is expanding its footprint in the financial services industry with the launch of Claude for Excel, an AI assistant integrated directly into Microsoft Excel. This integration allows financial analysts to interact with AI within their spreadsheets, enhancing productivity and accuracy in financial modeling and data analysis. Claude AI connects to real-time market data through partnerships with major financial data providers, offering enhanced transparency by explaining its spreadsheet actions at the cell level to build trust and ensure regulatory compliance.

The platform supports complex spreadsheet modifications, debugging, and new model creation, targeting banks, asset managers, and insurance companies. Alongside Excel, Claude models will be available in Microsoft Copilot Studio and Researcher agent, expanding AI integration across Microsoft’s enterprise AI ecosystem. Anthropic has also introduced pre-configured workflows, called Agent Skills, to automate common financial tasks such as discounted cash flow modeling and earnings analysis.

With significant clients including Bridgewater, Commonwealth Bank of Australia, and Norges Bank Investment Management, Anthropic reports massive productivity gains. Despite regulatory uncertainties and risks around AI errors, Anthropic emphasizes human oversight in AI use. The company's strategy highlights domain-specific AI enhanced with proprietary data access, positioning it as a strong competitor in the growing AI finance market.

MiniMax-M2, developed by the Chinese AI startup MiniMax, is gaining rapid recognition as the premier open-source large language model (LLM), particularly excelling in agentic tool usage—enabling autonomous application of external software like web search with minimal human input. Licensed under the enterprise-friendly MIT License, MiniMax-M2 is freely accessible for development and commercial use via platforms like Hugging Face, GitHub, and ModelScope, and supports OpenAI and Anthropic API standards. Independent evaluations by Artificial Analysis rank it highest among open-weight LLMs globally in intelligence and agentic capabilities, with performance rivaling top proprietary models like GPT-5 and Claude Sonnet 4.5. Its sparse Mixture-of-Experts architecture, with 10 billion active parameters from 230 billion total, enables efficient enterprise deployment on as few as 4 NVIDIA H100 GPUs, reducing costs and infrastructure demands. Benchmarks show MiniMax-M2’s strength in reasoning, coding, and complex tool-assisted workflows, making it ideal for developer automation, autonomous agents, and operational AI tasks. The model supports interleaved reasoning and structured tool calls, with comprehensive documentation for integration into enterprise systems. Priced competitively with an API cost of $0.30 per million input tokens and $1.20 per million output tokens, MiniMax-M2 offers significant cost-performance advantages. Backed by Alibaba and Tencent, MiniMax is a rising force in AI, known for innovations in video generation and extended context language models. MiniMax-M2 represents a significant milestone in open-weight AI, combining high-end technical capability with practical, enterprise-ready deployment options, reinforcing China's leadership in efficient, transparent AI development.