contentSlug

type

ContentSlug_ContentQuery

slug

ContentPageContentQuery

content

trendingContent

query ContentSlug_ContentQuery(
  $contentSlug: String!
  $type: [ContentType!]
) {
  ...ContentPageContentQuery_13QmVO
}

fragment CompanyDetails_Content on Content {
  __isContent: __typename
  companies {
    name
    description
    pageUrl
    id
  }
}

fragment ContentBasicInfo on Content {
  __isContent: __typename
  id
  type
  title
  lang
  createdAt
  ... on Video {
    liveTime
  }
  ... on Webcast {
    liveTime
  }
  companies {
    viewerIsFollowing
    id
    name
    pageUrl
    forum {
      threadUrl
    }
    defaultInstrument {
      symbol
      id
    }
  }
  ...PublishingDetails
  ...Reactions
}

fragment ContentMetadata on Content {
  __isContent: __typename
  type
  title
  lang
  pageUrl
  alternatePageUrls: translations {
    __typename
    pageUrl
    lang
  }
  imgUrl
  description
  createdAt
  changedAt
  isPremium
  authors {
    __typename
    name
    pageUrl
    id
  }
  companies {
    name
    pageUrl
    id
  }
  tags {
    name
    id
  }
  ... on Video {
    liveTime
    videoUrl
    videoEmbedUrl(location: "inderes")
  }
  ... on Webcast {
    liveTime
    videoUrl
    videoEmbedUrl(location: "inderes")
  }
}

fragment ContentPageContentQuery_13QmVO on Query {
  content(slug: $contentSlug, type: $type) {
    __typename
    id
    type
    title
    isPremium
    lang
    htmlBody
    summary
    companies {
      forum {
        threadUrl
        latestPosts {
          ...ForumPosts
        }
      }
      name
      id
      viewerIsFollowing
      pageUrl
      defaultInstrument {
        isin
        id
        symbol
        marketOpen
        marketClose
        marketplace {
          region
        }
        returns {
          total
        }
      }
    }
    tags {
      viewerIsFollowing
      id
      name
      pageUrl
    }
    createdAt
    ... on Video {
      liveTime
    }
    ... on Webcast {
      liveTime
    }
    slug
    ... on CompanyReport {
      reportUrl
      researchContentUrl
      researchContent {
        title
        lang
        pdfUrl
        relatedContent {
          __typename
          ...PublishingDetails
          ...Reactions
          id
        }
        content {
          title
          lang
          htmlBody
          publishedAt
          createdAt
          changedAt
          isOutdated
          isPremium
        }
      }
    }
    ... on ThirdPartyCompanyReport {
      reportUrl
    }
    ... on StockExchangeRelease {
      reportUrl
    }
    ...ContentBasicInfo
    ...ContentMetadata
    ...PriceAndRecommendation_Content
    ...CompanyDetails_Content
    ...ReportLink_Content
    ...KeyFigures_Content
  }
  ...TrendingContent_1jQ2Cn
}

fragment ForumPosts on ForumPost {
  text
  likesCount
  url
  author
  postedAt
  authorImgUrl
}

fragment KeyFigures_Content on Content {
  __isContent: __typename
  companies {
    defaultInstrument {
      estimateData {
        modelType
        reportingCurrency
      }
      snapshot {
        transaction {
          date
          actualized {
            year
            quarter
            revenue
            revenueGrowth
            ebitAdjusted
            ebitReported
            ebitda
            enterpriseValue
            epsAdjusted
            dividend
            dividendYield
            sharePrice
            nav
            navPerShare
            opexOfNav
            quarterlySharePrice
          }
          estimates {
            year
            quarter
            revenue
            revenueGrowth
            ebitAdjusted
            ebitReported
            ebitda
            enterpriseValue
            epsAdjusted
            dividend
            dividendYield
            sharePrice
            nav
            navPerShare
            opexOfNav
            quarterlySharePrice
          }
        }
      }
      id
    }
    id
  }
}

fragment PriceAndRecommendation_Content on Content {
  __isContent: __typename
  createdAt
  companies {
    name
    pageUrl
    defaultInstrument {
      priceData {
        tradeCurrency
      }
      snapshot {
        priceData {
          date
          lastPrice
          tradeCurrency
        }
        transaction {
          date
          recommendation
          targetPrice
        }
        prevTransaction {
          recommendation
        }
      }
      id
    }
    id
  }
}

fragment PublishingDetails on Content {
  __isContent: __typename
  publishers {
    name
    imgUrl
  }
  authors {
    __typename
    id
    name
    slug
    imgUrl
    jobTitle
  }
}

fragment Reactions on Content {
  __isContent: __typename
  id
  reactionGroups {
    reactionType
    reactionsCount
    viewerReacted
  }
}

fragment ReportLink_Content on Content {
  __isContent: __typename
  type
  isPremium
  ... on CompanyReport {
    reportUrl
  }
  ... on ThirdPartyCompanyReport {
    reportUrl
  }
  ... on StockExchangeRelease {
    reportUrl
  }
}

fragment TrendingContent_1jQ2Cn on Query {
  trendingContent(type: ARTICLE) {
    __typename
    id
    title
    lang
    pageUrl
    createdAt
    type
    imgUrl
    companies {
      squareLogoUrl
      name
      id
    }
    authors {
      __typename
      imgUrl
      id
    }
    publishers {
      imgUrl
    }
  }
}


__typename

title

isPremium

lang

htmlBody

summary

companies

createdAt

publishers

authors

reactionGroups

pageUrl

translations

imgUrl

description

changedAt

HAV Group operates in the energy and manufacturing industries. The company works with the development of energy solutions for the maritime industry. The product portfolio is broad and includes systems for hydrogen energy and ballast management. In addition to the main business, aftermarket service, support and maintenance are offered. Operations are held on a global level, with the largest presence in Scandinavia. The HAV Group was founded in 2021.

I don’t know anything about this specific company, and I sold my shares 1.5 years ago at the $140 level. That is why I dare to challenge this: If AI computing truly grows 10–100-fold over the next 5–10 years, why would Nvidia, of all companies, be the one whose market share collapses first?

Over the last few weeks and months, I have begun seriously considering liquidating my (disproportionately large) Nvidia position and diversifying more broadly across the tech, AI, and potentially even the first-wave application layer players—especially cyber and similar sectors. This is going to be a long post, which I’m writing primarily for myself to organize my train of thought. I know that, based on the numbers, Nvidia looks exceptionally cheap right now—cheaper than almost ever before in its history—but my confidence that Nvidia will be the completely dominant winner of the AI value chain is slowly fading. Saturation of Raw Compute Nvidia’s mega-rise has been based on the scaling of monolithic models in the zero-shot paradigm (Deepmind, 2022). This is now breaking down from at least three different directions: 1) physical constraints (electricity, construction, etc.), 2) data availability constraints (research papers from Epoc AI, Shumailov et al., etc., that are already 1.5–2 years old), and 3) economic constraints (swarms of small models seem to produce results just as good as giant parameter monoliths). Pre-training will certainly continue to evolve, but it’s starting to look to me like it can’t continue for much longer with a growing number of competing companies training competing base models—the economics just don’t support it. It also looks like various AGI “tinkerings” are fading, and the current goal is to get more out of what already exists by building harnesses around it, using smarter RAG (Retrieval-Augmented Generation), memory/context management, and optimization, etc.—more on these later. On top of everything, today’s best models are already being trained on competing chips (Google’s 8-series TPUs and Anthropic’s Amazon Trainium). Shifting Focus to Inference and Agentic Orchestration There has been a definitive breakthrough here in the last 6 months. Nvidia does not have the same level of competitive advantage in inference as it does in raw parallel computing. It is only just building its CPU offering, and various inference and test-time optimization tricks happen much further away from the hardware than training does. This is clearly visible in benchmarks. While zero-shot inference with monoliths has saturated at the latest with the o1/R1 reasoning models, ARC-AGIs and the like have started to rise as more has been extracted from inference (also demonstrating, in part, the saturation of raw compute for monolithic models). True, abstract generalization finally seems to happen when running agent swarms with suitable topologies that have access to the right tools. Research also seems to be heavily focused on extracting more from existing models by innovating around them, rather than trying to invent something entirely new within the model itself (which would require flexible compute hardware à la Nvidia). New RAG architectures, SML (Small Language Model) swarms / Multi-agent systems are now the “thing” that is fueling interest in new and easy-to-use orchestration tools (Copilot Studio, Codex, Claude code, LangGraph, etc.), even outside the AI scene and even here in Finland—based on my own empirical observations. Competitors’ Vertical Strategy I remember having written about this before. The challenge for Nvidia might not even come from AMD, Cerebras, or Huawei. It comes from the fact that hyperscalers like OpenAI and Anthropic have opened or are opening the game to their own vertical stack, where everything from the hardware to the application is self-developed. This is sensible if you can pull it off: it reduces strategic risks regarding at least one monopoly supplier (Nvidia) and, on the other hand, allows for the carving out of deep competitive advantages. Building chips is not easy, and the industry joke is that, at best, the third or fourth generation is usable. Google already has a vertical stack, rumors say Microsoft will soon have the third generation of MAIA, AWS/Anthropic are forming a stack together, etc. A kind of tipping point could therefore possibly be seen sooner rather than later. Synchronous Training vs. Asynchronous Agents (and the Breaking of the CUDA Moat) Monolithic model pre-training is, by nature, a tightly coupled and synchronous process. This physical requirement has been the ultimate defensive wall for Nvidia’s software moat, CUDA, as well as its rack/DC-scale strategy (Mellanox / Infiniband). Agentic orchestration, by contrast, is a largely asynchronous process. When an agent swarm solves a problem, the tasks are independent of each other. And because the tasks are broken down into small, asynchronous parts and are often run on smaller models (SLM), they fit into the memory of one or two chips. This asynchronicity destroys the moat from two directions:  The need for communication disappears: When a model doesn’t need to be split across thousands of chips, the need for Nvidia’s superior NCCL libraries and NVLink switches vanishes. At least in the current paradigm—though it is possible that advanced agent communication might eventually be implemented at the hardware level rather than through .md, .json, etc., documents and MCP/API calls. Once heavy synchronization is bypassed: The inference itself can be run on hardware-agnostic software. Open standards like OpenAI’s Triton compiler allow an AI model to be compiled directly onto, for example, AMD, Google, or AWS chips without a single line of CUDA code.  First Warnings of a CAPEX Cycle Plateau The overcapacity of Colossus and Meta is a small red flag for me. I’ve said before that Capex investments (together with TSMC’s demand outlook) are the most important signal for me to start thinking about exposure to AI “picks and shovels.” This isn’t a deal-breaker yet, primarily because the sales price of (Colossus) overcapacity was very high, indicating that there is indeed total demand for compute—it’s just no longer distributed as evenly as it was before.  I don’t believe that AI is currently in a bubble. On the contrary, following Jevons paradox, I think that demand will continue to grow as prices drop. But this inevitably shifts the focus of the value chain toward its narrowest points—and, in my view, these are no longer necessarily in Nvidia in the same way they have been for the last 2–3 years. I hope someone challenges the thoughts above

Here are the Mag7 valuation levels in light of the latest consensus forecasts. NVidia is the most affordable in terms of both P/E and PEG ratios.

“Undervaluation” is a term about which one could write books and discuss forever. I’ll take a stance on this by slightly sidestepping that tricky term. My personal view is that the market’s focus, which shifts through rotation, is currently elsewhere. The image below shows Nvidia’s volume at the bottom as dark bars. Trading volume for Nvidia’s stock has slowed down.  The stock exchange is an auction, and price is formed through supply and demand. In most cases, an increase or decrease in volume leads to a corresponding move in the share price. When there is interest in a stock, it is traded, and when rotation shifts the market’s focus elsewhere, volume drops and the price falls. In Nvidia’s case, we are near the ATH (all-time high), i.e., even the rotation hasn’t really pushed the price down. Strong fundamentals are supporting the price. Nvidia is a good example of how, over a two-year horizon, it is precisely these market rotations (“fashion trends”) that influence the share price the most. Nvidia’s trading volume has shifted, for example, to its subcontractors, such as memory companies. And in this way, Nvidia has become undervalued in the eyes of valuation metrics. It is interesting to consider whether this undervaluation will resolve into increasing demand, i.e., rising trading volumes. Personally, I believe the undervaluation will resolve within a couple of years.

According to the tweet, Nvidia is undervalued. The tweeter believes that AI will drive demand for chips for at least another decade, which reduces the significance of cyclical fluctuations in the industry.  It’s trading at just 16x 2027 earnings despite almost 32% annual growth expectation by 2028. Cheapest of the major designers.  https://x.com/oguzerkan/status/2075318004124930058

According to a research firm, the release of Nvidia’s next-generation Kyber server system is being pushed to 2028 due to manufacturing issues. The delay raises concerns about one of the company’s key competitive advantages—its rapid product release cadence—and could provide rivals such as AMD and Google with a rare opportunity to compete in high-end AI computing solutions. However, the current Rubin systems are still moving toward production as planned.  Key Points  The Kyber NVL144 has been pushed to 2028 due to difficulties in manufacturing a key circuit board: SemiAnalysis. The reported delay adds to concerns that Nvidia’s breakneck annual release cadence contested manufacturing limits. That delay could give rivals, such as AMD and Google, a rare technical opening at the high end of the market.   https://www.cnbc.com/2026/07/06/nvidia-kyber-rack-system-delays-manufacturing-taiwan-rubin-chips-.html

And it looks even more affordable when compared to the 2027 consensus estimates

HAV Group

HAV Group ASA (the "Company") has on 10 July 2026 sold 35.856 of its own shares at a price of NOK 13,39 per share (average closing price week 28 2026), in connection with the share purchase ...

<pre>HAV Group ASA (the "Company") has on 10 July 2026 sold 35.856 of its own shares at a price of NOK 13,39 per share (average closing price week 28 2026), in connection with the share purchase program for employees. The maximum subscription amount per employee was set at NOK 40,000, and a discount of NOK 1,500 has been given for each NOK 10,000 of subscription. After the sale of own shares, the Company owns zero treasury shares. This information is published in accordance with the Market Abuse Regulation (EU) 596/2014 Article 5, as supplemented by delegated Commission Regulation (EU) 2016/1052, and is subject to disclosure requirements in accordance with the Securities Trading Act § 5-12. s Trading Act § 5-12. \ </pre>

HAV: Sale of own shares

HAV: Sale of own shares