Big week next week, as the SpaceX IPO hype machine revs its engines (we even have a ticker: “SPCX”!). We’re preparing a massive report on this one and have spent much of the last year digging into the space sector (see our “2026: A Space Ontology” report from April to help understand the sector better). We also expect some big announcements from AI-arms-supplier Dell’s Dell Technology World.

You can find prior notes and the full research archive at https://epistrophy.beehiiv.com.

As always, I’m focused on three things:
1) Technology-driven change;
2) the latest in innovation and startup trends, and;
3) stock fraud.

The question is sometimes asked: what does the Oracle (ORCL:NYSE) datacenter buildout do for Oracle’s vaunted database business. The answer, increasingly, is to move every user interaction closer to the AI, giving Oracle software users a superior experience. 

That process took a big leap – little noticed – on May 12, when Oracle announced a new built-in MCP server for all of its database customers. MCP (“Model Context Protocol) acts like a restaurant waiter: taking a customer’s request, knowing which kitchen station can fulfill it, carry it quickly, securely and brings back just what was ordered. 

Oracle’s MCP update is a big deal because it gets the user request to the MCP securely, with clarity.  Just like that, every user will get exactly what they ordered – without sending anything back to the kitchen.

MCP is now  a production access layer for enterprises. The service s part of the no-cost OCI Database Tools service that's available from the OCI control plane, uses OAuth 2, supports streamable HTTPS, runs as a multi-user service and is offered at no added charge to OCI customers. This works with Oracle databases across deployment models, including Oracle Database@Azure, Database@AWS and Oracle Database@Google Cloud. 

The timing makes sense. MCP has advanced quickly over the past 12 months and vendors are now trying to convert early agent demos into governed enterprise workflows. Oracle first showed a local, single-user MCP server in June 2025. This release expands that into a centrally managed service aimed at organizations with dozens of groups and hundreds or thousands of employees. That shift matters because enterprise adoption depends less on whether a model can write SQL in a lab than on whether one hundred or one thousand or ten thousand workers can reach live systems without breaking access controls, auditability or operations.

Oracle’s core technical claim is identity propagation. A user authenticates through a chat interface, the identity flows through Oracle’s MCP layer and the database can record the actual end user in audit logs. Oracle said the system integrates with Oracle IDCS and federates to platforms such as Microsoft Entra ID. This level of escurity even  enables those chat interfaces to authenticate and authorize at that layer with the corporate SSO and have that propagated all the way down through the full stack. It’s built in managed identity.

In enterprise AI stacks, the hard part is rarely the model call. The harder problem is preserving who the user is, what the user can do and how those permissions survive movement across systems. Business users can sign in with a corporate identity, use a preferred AI client and carry that identity through to the database layer. Permissions can be controlled at the level of tool groups, individual reports and SQL visibility. One can imagine finance analysts, operators and developers all touching the same data through different agent workflows.

The second major element is Oracle’s treatment of natural language to SQL. Here the company was more realistic than many vendors. Oracle said production databases are often full of opaque names, incomplete metadata, missing foreign keys and schema conventions that make pure NL-to-SQL unreliable. Its answer is a report layer built from validated parameterized SQL queries. These reports become MCP tools. The model can discover them, match them to user intent and run them with explicit parameters.

That design addresses the trust problem directly. Oracle said business experts can That design addresses the trust problem directly. Oracle said business experts can define reports that cover roughly 80% of common questions. If that estimate holds, the model no longer needs to improvise SQL for most user requests. It can route intent to known query artifacts. 

Interestingly, reports do not cache result sets. They store query definitions. Each execution runs against live data. That matters in finance, operations and supply chain settings where yesterday’s answer can already be stale. Users with sufficient privileges can also inspect the SQL behind reports, giving developers and power users a starting point for customization. The model can use those stored queries as context, learning from joins, metric logic and schema patterns embedded in validated SQL. Repetition becomes the models friend, as it trains itself to improve with every query – the more you use it, the better it can get.

The platform’s support for asynchronous execution is another important detail. Oracle said many MCP interactions run into timeouts around 10 seconds, which is too short for warehouse-scale workloads. Its managed service can run queries asynchronously, materialize results into object storage and notify the user when processing is complete. That matters because enterprise analytics jobs often exceed the request-response assumptions built into current agent frameworks. Oracle also said the service includes built-in tools for ad hoc SQL, report discovery and execution, metadata retrieval for context engineering and asynchronous task handling. Customers can define custom tools as well, but administrators have to enable them explicitly. A newly created MCP server starts with no tools attached.

Operationally, Oracle said the service is native to OCI and uses OCI logs, metrics and analytics for observability rather than a separate MCP monitoring layer. It also said the service can scale through Oracle’s cloud footprint using the mechanics of the OCI control plane. On standards, Oracle emphasized that the service is an MCP server over HTTPS and should work with any compatible framework. The company cited collaboration with Microsoft’s Azure AI tooling and pointed to an open source Oracle MCP repository plus a separate Oracle Skills repository with more than 100 agent skills from Oracle database and Graal teams.

The economics are clear. Oracle is charging zero for the MCP server and monetizing the underlying database usage. That reduces adoption friction and fits Oracle’s broader platform strategy. If easier AI access increases database queries, active users and OCI consumption, Oracle captures the value at the infrastructure layer rather than the connector.

The most credible part of Oracle’s case is its realism. The company focused on identity, report-driven retrieval, audit trails, async execution and managed operations. Those are the constraints that decide whether an AI data access layer survives inside a large enterprise. The open question is execution quality: whether Oracle can make identity federation work cleanly across mixed environments, keep report catalogs maintainable at scale and extend the model into Oracle applications and mixed data estates. Still, the direction is clear. Oracle is trying to control the path between user intent and enterprise data. That is a strong position if it can deliver.

The Cerebras Systems (CBRS: NASDAQ) IPO arrives at a moment when artificial intelligence infrastructure has become strategically uncomfortable for its largest customers. Nvidia (NVDA: NASDAQ) controls much of the modern AI stack simultaneously: accelerators, networking, software tooling and increasingly the architecture of entire data centers. The company’s dominance helped create the AI boom. It also created an industry increasingly desperate for alternatives.

Cerebras is effectively selling investors on one core argument: Nvidia became too profitable for competitors not to emerge.

That tension runs through the entire AI economy. Nvidia’s gross margins climbed above 70% during the generative AI expansion, extraordinary economics for a hardware supplier operating at hyperscale volume. Customers accepted the pricing because OpenAI’s success transformed AI compute into an arms race. Hyperscalers needed GPUs immediately. Frontier labs needed training clusters immediately. Nvidia controlled both supply and the software ecosystem through CUDA, which increasingly functions like an operating system for AI development.

The result resembles a tax on participation in modern AI.

Cerebras built its architecture around attacking the bottlenecks Nvidia’s model depends on. Traditional GPU clusters distribute workloads across thousands of processors connected through increasingly complex networking fabrics. As models scale larger, communication overhead becomes one of the central technical problems in AI infrastructure. Data moves constantly between accelerators, memory pools and switches. Nvidia understood this years ago, which is why it acquired Mellanox and transformed itself into a systems company rather than merely a chip vendor.

Cerebras took the opposite path. Its Wafer-Scale Engine attempts to minimize communication overhead by keeping workloads local to one enormous processor. The latest WSE-3 architecture spans nearly an entire silicon wafer and contains roughly 4 trillion transistors, 900,000 AI-optimized cores and 44 gigabytes of on-chip SRAM memory. Instead of stitching together thousands of smaller processors, Cerebras concentrates computation inside one giant chip.

This is a big deal for inference workloads. The first phase of the AI boom rewarded whoever could train the largest models fastest. The next phase concerns serving those models economically across billions of daily queries. Token generation costs matter. Latency matters. Power consumption matters. A model that performs brilliantly but costs too much to operate becomes commercially fragile.

Cerebras positioned itself directly against those economics. The company repeatedly emphasized token throughput and low-latency inference demonstrations tied to OpenAI-style workloads. Its argument is straightforward: modern AI infrastructure spends too much energy and time moving data rather than generating intelligence.

Nvidia still dominates nearly every commercially important layer of the market. CUDA remains deeply entrenched across enterprise AI development. Developers optimize for Nvidia hardware because customers demand compatibility. Nvidia’s Blackwell systems also benefit from manufacturing scale competitors cannot easily match through Taiwan Semiconductor Manufacturing (TSM: NYSE) and advanced CoWoS packaging capacity.

Yet Nvidia’s extraordinary profitability changed customer incentives. Microsoft (MSFT: NASDAQ), Amazon.com (AMZN: NASDAQ), Alphabet (GOOG: NASDAQ) and Meta Platforms (META: NASDAQ) are spending tens of billions annually on AI infrastructure. At those levels, even modest pricing pressure creates enormous savings opportunities. A 15% reduction in infrastructure costs can justify financing entirely new accelerator ecosystems.

OpenAI sits at the center of this dynamic. The company helped create Nvidia’s modern dominance by proving transformer scaling worked commercially. ChatGPT triggered the hyperscale AI spending cycle that reshaped the semiconductor industry. OpenAI simultaneously became one of the strongest incentives for diversification away from Nvidia dependence. No frontier model developer wants a single supplier controlling pricing, supply availability and architectural direction simultaneously.

That explains why OpenAI reportedly works with both Nvidia and Cerebras. Optionality itself has become strategically valuable.

The broader competitive landscape reflects the same pressure. Advanced Micro Devices (AMD: NASDAQ) pushed aggressively into AI accelerators through its MI300 and MI350 families. Amazon built Trainium and Inferentia internally because hyperscale economics increasingly reward vertical integration. Google spent years developing Tensor Processing Units for similar reasons. Groq emphasized deterministic inference throughput. The still-private Positron emerged with a more openly confrontational argument that GPU-centric AI infrastructure has become economically bloated for inference-heavy workloads.

None individually threaten Nvidia’s dominance near term. Collectively they matter because they weaken the assumption that one company permanently controls AI compute economics.

Cerebras nevertheless faces formidable risks entering public markets. Wafer-scale computing introduces manufacturing challenges conventional chiplet architectures largely avoid. Yield management becomes more difficult when nearly an entire wafer functions as one processor. Nvidia meanwhile benefits from modular scaling across accelerators, networking and memory systems. More importantly, Nvidia possesses software inertia competitors struggle to overcome. CUDA’s ecosystem advantage extends beyond code compatibility into tooling, developer training and enterprise deployment workflows.

That creates the central tension inside the Cerebras IPO. The company is asking investors to believe AI infrastructure needs architectural change while acknowledging the dominant software environment belongs to Nvidia.

Power constraints may ultimately determine whether Cerebras succeeds. AI infrastructure increasingly collides with electrical and cooling limitations. Hyperscalers now discuss gigawatt-scale campuses. Utilities struggle to meet projected data center demand growth. Under those conditions, architecture efficiency becomes strategically important rather than merely incremental. A materially faster or lower-power inference system changes deployment economics, utility planning and data center design simultaneously.

The semiconductor industry rarely tolerates prolonged excess extraction. Intel (INTC: NASDAQ) once controlled CPU economics with similar confidence. Cisco Systems (CSCO: NASDAQ) once dominated networking infrastructure with similarly formidable ecosystem advantages. Each eventually created enough profit opportunity for competitors to emerge around them.

Nvidia remains the central power in AI infrastructure. The company built the modern AI stack and still controls the industry’s strongest ecosystem advantages. Yet dominance at this scale changes customer behavior. Buyers search for leverage. Suppliers search for architectural asymmetries. Venture capital finances escape routes.

The Cerebras IPO is ultimately a wager that the AI industry no longer wants a single company collecting tolls across the entire compute economy.

On Tuesday NewsNation had me on to discuss Anthropic’s “peace talks” with the White House. And in a brief swing through Manhattan (and yes, I’ll take Manhattan) I was able to join Julie Hyman on set at Yahoo! Finance for a good 15 min to talk about the Cerebras IPO, NVIDIA earnings AND what to expect from the SpaceX filing. I had a surprise visit with CNBC’s terrific Sara Eisen. Nicole Petallides hosted me at the New York Stock Exchange. And finally I had a great visit with Collin McShane at NewNation. Great conversations all!

I’ll be back in our San Francisco office at the Ferry Building all week digging deeper and deeper into the IPO mania. Call or, heck, come visit!

Written reports are available to clients, with video summaries on YouTube, and of course our popular summaries of the summaries on Instagram, TikTok, and YouTube Shorts.