The Answer Engine AI News · Sydney

Anthropic, the AI company behind Claude (the chatbot that competes with ChatGPT), is preparing for a potential IPO, and its early investors are already looking at significant returns on paper. The company has attracted backing from major players including Google, Salesforce, and various venture capital firms. These investors got in early when Anthropic was valued much lower than its current estimated valuation in the billions. While the exact figures aren't public, the gap between what early backers paid and what the company is now worth represents substantial gains—even before shares trade publicly. This reflects the broader AI investment boom, where companies building large language models have seen valuations skyrocket. For context, Anthropic was founded in 2021 by former OpenAI executives and has raised multiple funding rounds as it developed Claude. The pending IPO would give the public a chance to invest, but early institutional and venture investors are the ones positioned to see the biggest multiples on their original stakes.

inDrive, a ride-hailing platform that competes with Uber, has announced plans to test self-driving taxis in Astana, Kazakhstan's capital city, starting in the first half of 2027. The company is partnering with Pony.ai, a Chinese autonomous vehicle technology firm, and chip giant Nvidia to make this happen. This marks a significant expansion of robotaxi technology beyond the usual testing grounds in the United States and China. inDrive operates differently from Uber by allowing riders and drivers to negotiate fares directly, and it has built a presence in emerging markets across Central Asia, Latin America, and Africa. The planned tests signal that autonomous vehicle technology is spreading to new regions, with Kazakhstan becoming an unexpected early adopter. While companies like Waymo and Cruise have focused on major US cities, and Chinese firms have tested extensively at home, inDrive's move shows that the race for self-driving dominance is becoming truly global, with emerging markets now entering the picture as testing and deployment locations.

Everyone talks about the cost of GPUs and power to run AI models, but there's a less visible expense piling up: data storage. Every time someone uses an AI tool—whether it's a text prompt, image generation, or video query—the system creates backend data that needs to be stored. With billions of daily interactions across ChatGPT, Perplexity, Midjourney and similar tools, this data storage requirement is growing fast. The issue affects companies running AI services and enterprises deploying AI internally. Unlike training data, which is a one-time cost, inference data is ongoing—every conversation, every image created, every refinement generates more files that need keeping for compliance, quality improvement, or legal reasons. This storage burden isn't just about hard drives; it includes backup systems, retrieval infrastructure, and the energy to keep it all running. For businesses adopting AI tools or building AI features into their products, storage costs could quietly become a significant line item that wasn't in the original budget. It's a reminder that AI's total cost of ownership extends well beyond the headline GPU prices.

Researchers have measured, for the first time, how much time and computing resources AI agents waste just introducing themselves before they can do useful work. When AI agents need to communicate—whether that's ChatGPT calling a weather service or a business assistant checking inventory—they first complete a "handshake": finding each other, discovering capabilities, and establishing a session. The study benchmarked three major protocols (MCP, A2A, and ACP) and found dramatic differences. Legacy MCP handshakes require three back-and-forth exchanges and take 230 milliseconds under typical network conditions; A2A needs just one round trip and 61 milliseconds. The cost multiplies when large language models read a counterpart's capability list—each capability costs 24-107 tokens to process, meaning an agent advertising 200 capabilities becomes expensive to work with. The researchers tested each protocol 50 times across realistic scenarios including cold starts, repeat connections, and degraded conditions. Round-trip count emerged as the dominant cost factor, though token processing for capability discovery adds significant expense when agents offer many services.

Researchers have compared two ways AI agents connect to external systems like databases, files, and cloud services. The first method, CLI (Command Line Interface), lets AI agents run shell commands directly—simple and fast for single users working locally. The second, MCP (Model Context Protocol), uses a structured JSON-RPC approach that provides governed, typed access to tools. The study evaluated both across 14 dimensions including security, efficiency, and governance. Key findings: CLI wins for efficiency in simple, local workflows. MCP handles requirements CLI can't meet—fine-grained access control, credential isolation, structured audits, and multi-user environments. The researchers found that newer "progressive disclosure" techniques solve MCP's traditional weakness (eating up too much context window), making it nearly as token-efficient as CLI while keeping its security and governance advantages. The paper recommends production systems use a hybrid approach, combining both methods depending on the use case.

Marie Haynes argues that while businesses are focused on Generative Engine Optimisation (GEO), the real shift is toward what she calls the "agentic web" — where AI assistants will actively make decisions and take actions on behalf of users. She points to emerging technologies like Universal Cart, UCP (Universal Checkout Protocol), WebMCP, and intelligent eyewear as examples of this trend. Instead of simply answering questions, these AI agents will compare products, make purchases, and complete transactions autonomously. Haynes suggests most businesses are missing this bigger picture by obsessing over how their content appears in ChatGPT or Perplexity results. The implication is that companies need to prepare for a future where AI doesn't just recommend their business — it actually transacts with it. This means ensuring your business systems can integrate with these emerging protocols and agent-based technologies, rather than just optimising content for AI-generated answers.

OpenAI has lifted the lid on how advertising works inside ChatGPT, revealing a system that's more complex than traditional search ads. The platform uses a custom large language model (LLM) as an "orchestrator" to understand multi-turn conversations – meaning it tracks what you've asked across several questions, not just your latest prompt. This orchestrator nominates which ads should enter the auction based on conversational context and relevance. From there, advertisers' bids compete for placement, but there's a catch: ads that don't meet response time deadlines get dropped entirely, even if they won the bid. This means speed matters as much as money. The system is designed to balance commercial intent with user experience, attempting to serve ads that feel natural within an ongoing conversation rather than disrupting it. For businesses considering ChatGPT advertising, this means campaign success will depend on three factors: bidding competitively, ensuring your ad creative and targeting are genuinely relevant to conversational context, and maintaining technical infrastructure that can respond to OpenAI's system quickly enough to meet placement deadlines.

Google's CEO Sundar Pichai has outlined how AI agents will access website content in the near future. He expects agents to interact with sites in two distinct ways: by browsing pages like traditional users do, and through structured APIs (application programming interfaces). According to Pichai, API-based protocols will ultimately scale better and be more efficient, but they'll take significantly longer to develop and implement across the web. In the shorter term, AI agents will continue accessing sites by crawling and reading pages much like search engines currently do. This dual approach suggests Google is preparing for a transition period where both methods coexist. For website owners, this means the familiar world of page-based access isn't disappearing overnight, but there's a clear signal that structured data delivery through APIs represents the long-term future. The timeline for this shift remains uncertain, giving businesses time to adapt their digital infrastructure without immediate panic.

New research from IDC reveals a significant gap in how much money AI leaders spend on modernisation compared to other businesses. Companies getting stronger results from AI are investing 53% of their budgets in modernisation efforts, while the rest are spending just 19%. The research suggests this investment gap directly correlates with AI outcomes – those who spend more on updating their systems and infrastructure are seeing better returns from their AI initiatives. The findings highlight that successful AI adoption isn't just about buying AI tools; it requires proper foundation work. Modernisation includes updating legacy systems, improving data infrastructure, and ensuring the business has the technical capability to support AI effectively. For Australian small businesses considering AI, the message is clear: budget allocation matters. Companies trying to bolt AI onto outdated systems without proper modernisation investment are falling behind competitors who take a more comprehensive approach. The research indicates that viewing AI as an isolated technology purchase, rather than part of a broader modernisation strategy, is a key differentiator between businesses that succeed with AI and those that struggle.